Intelligent consultation interaction method based on CRM platform

Abstract

The invention relates to the technical field of intelligent interaction, and particularly discloses an intelligent consultation interaction method based on a CRM platform, and the method comprises the following steps: S1, obtaining a text sent by a user, obtaining the semanteme of the text, and determining a target semanteme; obtaining historical semantics, and determining undetermined semantics in the historical semantics; s2, obtaining a reply A of a text corresponding to the undetermined semantic meaning, and obtaining a text B1 based on a time point when the reply A ends; obtaining a semantic analysis result of the text B1, and determining a standard semantic based on the semantic analysis result; and S3, clustering the standard semantics to obtain clusters, obtaining representative semantics in the cluster with the maximum density, and replying to the user based on the representative semantics. The method can effectively reduce the burden of manual customer service and improve the overall service efficiency and quality.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent interaction, and particularly relates to an intelligent consultation interaction method based on a CRM platform. Background Art

[0002] The CRM (Customer Relationship Management) platform is a software system designed to help enterprises effectively manage their interactions and relationships with customers. The consultation interaction method is that enterprises use the CRM platform and artificial intelligence technology to enable customers to quickly obtain intelligent answers when consulting. For example, when a customer consults a question through online chat, the system will automatically identify and understand the customer's needs and reply according to the pre-set answers; if the question is more complex, it will be transferred to a human customer service for processing.

[0003] Although this method reduces the workload of human customer service to a certain extent and improves the response speed, there are still some limitations. The system relies on pre-set rules and answers, and its flexibility and adaptability will be restricted when facing the changing market environment and diverse customer consultations. There may be many questions that cannot be replied and need to be handed over to human customer service for processing, which will increase the workload of human customer service. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent consultation interaction method based on a CRM platform to solve the following technical problems: The system relies on pre-set rules and answers, and its flexibility and adaptability will be restricted when facing the changing market environment and diverse customer consultations. There may be many questions that cannot be replied and need to be handed over to human customer service for processing, which will increase the workload of human customer service.

[0005] The purpose of the present invention can be achieved through the following technical solutions: An intelligent consultation interaction method based on a CRM platform, S1: Obtain the text sent by the user, obtain the semantics of the text. If there is no reply corresponding to the semantics, mark the semantics as the target semantics; Obtain the semantics of the text sent by the user historically, denoted as historical semantics, and mark the historical semantics with a similarity degree greater than a preset value to the target semantics as the semantics to be determined; S2: Obtain the reply A of the text corresponding to the semantics to be determined, obtain the time point t1 when the reply A ends, and record the text sent by the user with the shortest time interval after the time point t1 and to the time point t1 as the text B1; Obtain the semantic analysis result of the text B1, and the semantic analysis result includes positive, neutral, and negative. If the semantic analysis result is positive or neutral, obtain the semantics of the text B1, denoted as the standard semantics; S3: Cluster the standard semantics to obtain clusters, and obtain the representative semantics in the cluster with the highest density. The total similarity between the representative semantics and the remaining standard semantics in the cluster is the highest. Reply to the user based on the representative semantics.

[0006] As a further solution of the present invention: in step S2, during the process of determining the standard semantics, if the semantic analysis result is negative, perform the following steps: Obtain the time point t2 when text B1 is sent, and record the text that is not sent by the user and has the shortest time interval after time point t2 and with respect to time point t2 as text B2; Obtain whether the semantic analysis result of the judgment text corresponding to text B2 is negative. If not, obtain the semantics of text B2 and record it as the standard semantics.

[0007] As a further solution of the present invention: in step S1, the process of obtaining the semantics of the text specifically includes: Preprocess the text. The preprocessing includes word segmentation, stop word removal, and part-of-speech tagging. Input the preprocessed text into a pre-trained language model, and output a semantic vector. The semantic vector contains the context and implicit information in the text, and is recorded as the semantics of the text.

[0008] As a further solution of the present invention: in step S1, the process of determining whether there is a reply corresponding to the semantics specifically includes: Obtain a preset reply table. The reply table stores replies and corresponding semantic vectors, and mark the semantic vectors in the reply table as comparison vectors; Obtain the cosine value of the included angle between the semantic vector of the text and the comparison vector. If the maximum cosine value of the included angle is less than a preset threshold, it is determined that there is no reply corresponding to the semantics.

[0009] As a further solution of the present invention: in step S2, the process of obtaining the semantic analysis result of text B1 specifically includes: Establish a database, and the database stores texts with annotated semantic analysis results; Establish a semantic analysis model based on a deep learning model, train and verify the semantic analysis model based on the database, and input text B1 into the trained and verified semantic analysis model to output the semantic analysis result of text B1.

[0010] As a further solution of the present invention: in step S3, the cosine value of the included angle between any two semantics in the cluster is greater than a preset distance threshold.

[0011] As a further solution of the present invention: in the step S3, if there are two or more clustering clusters with the same and maximum density, they are sent to a pre-set reviewer for review.

[0012] As a further solution of the present invention: in the step S3, the process of replying based on the representative language intention user specifically includes: using the representative language meaning as a condition to input into a natural language generation model, and the natural language generation model generates a reply text that conforms to the grammatical structure and semantic logic.

[0013] The beneficial effects of the present invention: compared with the prior art: 1) Through the comprehensive application of text preprocessing, semantic vector extraction, and clustering analysis and other links, this solution can capture the context information and potential meaning in the user text more deeply; Context and implicit information capture: With the help of the pre-trained language model, the system can not only identify explicit keywords, but also perceive context associations and potential semantics, making the reply more in line with the user's real needs; Diversified scenario adaptation: In the changing market environment and customer needs, the system can adapt to newly emerging problems and quickly judge whether there are corresponding answers, reducing a large number of rule-based constraints.

[0014] 2) In the present invention, based on the in-depth semantic analysis of the text, combined with clustering technology and natural language generation, it breaks the limitations of traditional template or rule-based replies; Dynamic clustering and rapid expansion: Through the clustering analysis of historical semantics and new questions, it can quickly discover and integrate similar questions, form new replies or accurately match on the basis of existing replies, and flexibly respond to different scenarios; Automatic semantic determination: Multi-dimensional emotion or attitude judgments such as positive, neutral, and negative help the system to seek information other than that sent by the user for secondary confirmation in a timely manner when facing negative emotion scenarios, thus reducing misjudgments; 3) Compared with the traditional mode of "transferring to manual when a problem cannot be matched", this solution uses an intelligent method to screen, learn, and update content, greatly reducing the repetitive workload of artificial customer service, and can reduce ineffective transfers: Because the system can automatically identify and reply to common questions more, complex questions do not have to be immediately transferred to the manual, thus saving customer service resources and improving manpower efficiency; Assisting customer service decision-making: For situations that cannot be automatically replied, the system can provide similar questions and corresponding replies after clustering, assisting the artificial customer service to quickly find historical experience and give a more targeted reply.

[0015] 4) Realize the automatic precipitation and iteration of knowledge Through continuous collection and analysis of text semantics and user feedback, this solution can achieve adaptive updates to the knowledge base; iterative optimization: in the iterative process of clustering and semantic analysis, continuously accumulate "standard semantics" and existing reply answers. As the interaction volume between the system and customers increases, the coverage and accuracy of intelligent replies also increase; traceable historical records: the process of marking historical sent texts and performing similarity retrieval enables the system to retain a large amount of interaction data, achieve knowledge precipitation, and quickly match or generate appropriate answers when new problems appear subsequently; In summary, through the mutual cooperation of multiple technical means such as semantic vector extraction, sentiment analysis, dynamic clustering, and natural language generation, the present invention realizes an intelligent, flexible, and iteratively updatable customer consultation interaction process, which not only reduces the pressure on artificial customer service but also maintains high adaptability and service quality in a constantly changing market environment, thus truly achieving the goal of "making customer service more efficient and customers more satisfied". BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below with reference to the accompanying drawings.

[0017] Figure 1 It is a schematic flowchart of an intelligent consultation interaction method based on a CRM platform according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Please refer to Figure 1 As shown, the present invention is an intelligent consultation interaction method based on a CRM platform, including the following steps: S1: Obtain the text sent by the user, obtain the semantics of the text. If there is no reply corresponding to the semantics, mark the semantics as the target semantics; In a preferred embodiment of the present invention, the process of obtaining the semantics of the text specifically includes: Preprocess the text. The preprocessing includes word segmentation, stop word removal, and part-of-speech tagging. Input the preprocessed text into a pre-trained language model to output a semantic vector, which contains the context and implicit information in the text and is recorded as the semantics of the text.

[0020] It is understandable that the system first preprocesses the text sent by the user, performs word segmentation, removes stop words, and conducts part-of-speech tagging. Then, the processed text is input into a pre-trained language model to obtain semantic vectors, enabling the system to accurately capture the user's intent and quickly screen. By deeply representing the semantic vectors, it can capture context and implicit information, significantly enhancing the ability to understand questions and providing a solid foundation for subsequent automatic replies. Exemplarily, when a user inputs a text on the CRM platform, the system first fully obtains it and performs corresponding encoding or escape processing (such as converting special characters into recognizable forms) to ensure that subsequent operations can proceed normally. Common Chinese word segmentation algorithms or tools can be used (such as methods based on dictionary matching or using deep learning models for word segmentation) to split the entire sentence into several tokens. For example, for the sentence "I want to know about the refund process", the system will segment it into tokens such as "I, want, know, about, refund, process". After word segmentation, the system uses a pre-compiled "stop word" list to filter these tokens, removing common words or symbols with limited semantic contribution, such as "de", "di", "le", "a", and some meaningless punctuation marks. Based on word segmentation and stop word removal, the system can also use existing natural language processing tools or models to perform part-of-speech tagging (such as nouns, verbs, adverbs, etc.) for each remaining token, further enriching the feature information of these tokens. Convert the series of tokens (and their part-of-speech tagging results) obtained after the above processing into an input form acceptable to the model (such as mapping the tokens into corresponding word vector sequences) and input them into the pre-trained language model. Here, large-scale pre-trained models such as BERT, GPT, or others can be used. These models have learned deep semantic and context information from extensive text data and can thus encode the input token sequence more accurately. After the pre-trained language model calculates the input token sequence, it outputs a vector representation (Embedding). To aggregate the overall semantic expression of the entire sentence or text segment, usually one of the following methods is adopted: 1) Take the vector corresponding to the [CLS] (classification marker) output by the model as the semantic representation of the entire sentence; 2) Use average pooling or max pooling methods to pool the vectors corresponding to all tokens to obtain an overall vector representation; 3) Combine the attention mechanism to weight and merge the vectors of each token to obtain a more contextually relevant information representation.

[0021] In a preferred case of this embodiment, the process of determining whether there is a reply corresponding to the semantics specifically includes: Obtain a preset reply table, where the reply table stores replies and corresponding semantic vectors, and mark the semantic vectors in the reply table as comparison vectors. Obtain the cosine value of the included angle between the semantic vector and the comparison vector of the text. If the maximum cosine value of the included angle is less than the preset threshold, it is determined that there is no reply corresponding to the semantics.

[0022] It should be noted that all predefined reply contents (i.e., answer texts) and their corresponding semantic vectors are loaded from the pre-constructed "reply table", and these semantic vectors are uniformly marked as comparison vectors. Subsequently, for the text to be detected, the system first obtains its semantic vector through the same or similar method as the previous step (such as using a pre-trained language model, etc.), and then compares the similarity with all comparison vectors in turn. The system will calculate the cosine similarity between the current text and all comparison vectors one by one, and then obtain its maximum value. If this maximum cosine similarity is still lower than the threshold set by the system in advance (usually between 0 and 1, such as 0.8 or 0.85), it means that the semantics of the current text is not close enough to any existing reply in the preset reply table, and thus it cannot be determined that there is a corresponding automated reply, so it is determined that "there is no reply corresponding to the semantics".

[0023] Obtain the semantics of the text sent by the user historically, denoted as historical semantics, and mark the historical semantics whose similarity degree with the target semantics is greater than the preset value as pending semantics; S2: Obtain the reply A of the text corresponding to the pending semantics, obtain the time point t1 when the reply A ends, and record the text sent by the user with the shortest time interval after the time point t1 and between the time point t1 as text B1; Obtain the semantic analysis result of the text B1, where the semantic analysis result includes positive, neutral, and negative. If the semantic analysis result is positive or neutral, obtain the semantics of the text B1, denoted as standard semantics; It should be noted that in order to determine the user's feedback attitude after a certain historical interaction semantics (i.e., pending semantics), it is necessary to first find the system reply A corresponding to this semantics and locate the time point t1 when it ends completely. Here, "ending" can be understood as that when the system sends several consecutive replies for this semantics, the last reply being successfully sent is regarded as the reply end time point. After that, the system will retrieve all texts sent by the user after t1 from the conversation record and find the user text closest to t1 in chronological order, and record it as text B1. This can ensure that B1 maximally reflects the user's first-time feedback after receiving reply A and has strong relevance; In a preferred embodiment of the present invention, in step S2, during the process of determining the standard semantics, if the semantic analysis result is negative, perform the following steps: Obtain the time point t2 when the text B1 is sent, and record the text that is not sent by the user and has the shortest time interval after the time point t2 and between the time point t2 as text B2; Obtain whether the semantic analysis result of the judgment text corresponding to the text B2 is negative. If not, obtain the semantics of the text B2, denoted as the standard semantics.

[0024] It should be noted that when the user initially expresses negative emotions towards a certain reply, the system does not immediately regard this negative text as the "standard semantics" that can be incorporated into subsequent processing. Instead, it continuously pays attention to the user's subsequent feedback. If the user turns to express approval or acceptance of the previous reply in a subsequent text (that is, the semantic analysis result is positive or neutral), it indicates that the user ultimately has a positive attitude towards the reply. At this time, the system will incorporate the semantic information corresponding to this text expressing approval into the "standard semantics". This can avoid misjudgment due to the user's momentary emotional negative expression and ensure that the knowledge base and subsequent automatic reply strategies are based on the content finally approved by the user, so as to more accurately reflect the user's true needs and attitudes. It can be understood that the semantic analysis result is used to reflect whether the user is satisfied with the current answer. Positive and neutral indicate that the user is satisfied with the current answer, while negative indicates that the user is not satisfied with the current answer. In this case, the user usually makes another inquiry. In a preferred embodiment of the present invention, in the step S2, the process of obtaining the semantic analysis result of the text B1 specifically includes: Establish a database, and the database stores texts with marked semantic analysis results. Establish a semantic analysis model based on a deep learning model, train and verify the semantic analysis model based on the database, input the text B1 into the trained and verified semantic analysis model, and output the semantic analysis result of the text B1 S3: Cluster the standard semantics to obtain a clustering cluster, obtain the representative semantics in the clustering cluster with the largest density, and the total similarity degree between the representative semantics and the remaining standard semantics in the clustering cluster is the highest. Reply to the user based on the representative semantics. In a preferred case of this embodiment, in the step S3, the cosine value of the angle between any two semantics in the clustering cluster is greater than a preset distance threshold.

[0025] It is worth noting that in order to ensure that the standard semantics entering the same clustering cluster have sufficient similarity, the system will first calculate the cosine value of the angle between each standard semantics and the existing semantics in the clustering. Only when this similarity is greater than the pre-set distance threshold will it be incorporated into the same clustering cluster. The setting of this threshold is generally based on a comprehensive consideration of the actual business scenario and data distribution. It is necessary to ensure a high degree of consistency of the semantics within the cluster and also avoid merging relatively irrelevant semantics together, resulting in a decline in clustering quality. In another preferred embodiment of the present invention, in step S3, if there are two or more clustering clusters with the same and maximum density, they are sent to a pre-set reviewer for review.

[0026] In a preferred embodiment of the present invention, in step S3, the process of replying based on the representative language intention user specifically includes: taking the representative language meaning as a condition and inputting it into a natural language generation model, and the natural language generation model generates a reply text that conforms to the grammatical structure and semantic logic.

[0027] It should be noted that the system has pre-integrated a trained natural language generation model, such as a pre-trained large model based on the Transformer architecture (such as GPT, etc.). This model can learn the basic grammatical rules and context logic of the language from a large amount of text data. At this time, the system will load the corresponding generation model or its corresponding fine-tuned version according to the current conversation scenario or the business field it is in, so as to ensure that the reply is more in line with the specific business requirements; the obtained representative language meaning can be regarded as the "input condition" or "prompt information (Prompt)" of the model. It usually includes the common core intentions or key points of the questions of the users in this clustering cluster. Map or convert the representative language meaning vector into natural language fragments (or embed corresponding keywords and short sentences in the Prompt), and inject them into the generation model together with the existing context information; after receiving the representative language meaning condition, the natural language generation model will automatically organize the language, conduct multiple rounds of prediction and sampling to generate sentences or paragraphs that conform to the grammatical structure and can effectively answer the user's needs semantically; if higher accuracy is required, certain sampling strategies and constraint rules can be set, such as restricting the length of the generated text, controlling the use of specific vocabulary or styles, so that the generated content is more in line with the enterprise's reply specifications; after obtaining the initially generated reply text, the system can then conduct basic verification on it, such as: keyword check: check whether the generated text contains sensitive words, privacy information or misleading expressions that the enterprise requires to avoid. Semantic consistency judgment: compare it with the representative language meaning again to confirm that the generated reply does not deviate significantly from the user's intention; if the result passes the verification, the system will finally present it to the user; if a large deviation or non-compliance is detected, it can automatically trigger re-generation or transfer to manual review to ensure the accuracy and compliance of the reply.

[0028] The above has described an embodiment of the present invention in detail, but the content described is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of application of the present invention should still fall within the scope covered by the patent of the present invention.

Claims

1. An intelligent consultation interaction method based on a CRM platform, characterized in that, It includes the following steps: S1: Obtain the text sent by the user, obtain the semantics of the text. If there is no corresponding reply for the semantics, mark the semantics as the target semantics; Obtain the semantics of the text sent by the user historically, denoted as the historical semantics, and mark the historical semantics with a similarity degree greater than a preset value to the target semantics as the semantics to be determined; S2: Obtain the reply A of the text corresponding to the semantics to be determined, obtain the time point t1 when the reply A ends, and record the text sent by the user with the shortest time interval after the time point t1 and to the time point t1 as the text B1; Obtain the semantic analysis result of the text B1. The semantic analysis result includes positive, neutral, and negative. If the semantic analysis result is positive or neutral, obtain the semantics of the text B1, denoted as the standard semantics; S3: Cluster the standard semantics to obtain a cluster. Obtain the representative semantics in the cluster with the largest density. The total similarity degree between the representative semantics and the remaining standard semantics in the cluster is the highest, and reply to the user based on the representative semantics.

2. The intelligent consultation interaction method based on a CRM platform according to claim 1, wherein In the step S2, during the process of determining the standard semantics, if the semantic analysis result is negative, perform the following steps: Obtain the time point t2 when the text B1 is sent, and record the text that is not sent by the user with the shortest time interval after the time point t2 and to the time point t2 as the text B2; Obtain whether the semantic analysis result of the judgment text corresponding to the text B2 is negative. If not, obtain the semantics of the text B2, denoted as the standard semantics.

3. An intelligent consultation interaction method based on a CRM platform according to claim 1, characterized in that In the step S1, the process of obtaining the semantics of the text specifically includes: Preprocess the text. The preprocessing includes word segmentation, stop word removal, and part-of-speech tagging. Input the preprocessed text into a pre-trained language model, and output a semantic vector. The semantic vector contains the context and implicit information in the text, denoted as the semantics of the text.

4. An intelligent consultation interaction method based on a CRM platform according to claim 3, characterized in that, In the step S1, the process of determining whether there is a corresponding reply for the semantics specifically includes: Obtain a preset reply table. The reply table stores replies and corresponding semantic vectors, and mark the semantic vectors in the reply table as comparison vectors; Obtain the cosine value of the included angle between the semantic vector of the text and the comparison vector. If the maximum cosine value of the included angle is less than a preset threshold, it is determined that there is no corresponding reply for the semantics.

5. The intelligent consultation interaction method based on the CRM platform according to claim 1, wherein In the step S2, the process of obtaining the semantic analysis result of the text B1 specifically includes: Establish a database. The database stores texts with marked semantic analysis results; Establish a semantic analysis model based on a deep learning model, train and verify the semantic analysis model based on the database, and input the text B1 into the trained and verified semantic analysis model to output the semantic analysis result of the text B1.

6. An intelligent consultation interaction method based on a CRM platform according to claim 4, characterized in that, In the step S3, the cosine value of the included angle between any two semantics in the cluster is greater than a preset distance threshold.

7. An intelligent consultation interaction method based on a CRM platform according to claim 1, characterized in that, In the step S3, if there are two or more clusters with the same and maximum density, send them to the pre-set reviewers for review.

8. An intelligent consultation interaction method based on a CRM platform according to claim 1, characterized in that In the step S3, the process of replying to the user based on the representative language intention specifically includes: using the representative language meaning as a condition to input into the natural language generation model, and generating a reply text that conforms to the grammatical structure and semantic logic by the natural language generation model.

Images