Intelligent customer service dialogue strategy optimization method based on deep learning

Abstract

The invention relates to the technical field of intelligent customer service, in particular to an intelligent customer service dialogue strategy optimization method based on deep learning, and the method comprises the steps: S1, obtaining first text information of a previous round of user text, and obtaining a text reply library corresponding to the previous round of user text according to the first text information; s2, performing emotion matching on the recognized keywords according to the text reply library corresponding to each round of user text, and determining the emotion tendency and intention of the user text; s3, based on the emotional tendency and intention of the user text, generating a reply text of the customer service, and determining a predicted emotional feature and a predicted reply feature of the reply text; s4, acquiring feedback information of the user, and matching the feedback information of the user with the reply text of the customer service to obtain a matched target sequence; obtaining a customer service dialogue strategy according to the matched target sequence; and the accuracy and efficiency of customer service dialogues are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent customer service, and specifically to a method for optimizing intelligent customer service dialogue strategies based on deep learning. Background Art

[0002] With the rapid popularization and development of online e-commerce, the work of e-commerce customer service staff is becoming increasingly busy, with high workload and repetitive and boring content. The professional skills of customer service staff are also uneven, and the turnover rate is high. With the development of the Internet, artificial intelligence technology has been widely used in finance, e-commerce, telecommunications and other industries, replacing some of the work of traditional manual customer service to a large extent, effectively alleviating the pressure of manual customer service.

[0003] For example, Chinese patent publication No. CN112507094A discloses a customer service robot dialogue method based on reinforcement learning and its related components, the method comprising: obtaining the first text information input by the user in the previous round, and selecting the corresponding action strategy for reply; obtaining the second text information of the user's feedback on the action strategy, using a pre-trained emotion classification model to perform emotion classification on the second text information, obtaining the user's emotion, and calculating the total reply value score according to the maximum expectation algorithm; converting the text information input by the user into a text vector, and inputting it into a multi-layer neural network for reinforcement learning to obtain the probability of selecting the corresponding action strategy, optimizing the multi-layer neural network according to the total reply value score and the probability of the action strategy, and obtaining the customer service robot reply model.

[0004] In the prior art, the reply text is evaluated to determine the value of each reply and the action strategy. However, when there are multiple input words, this reply method results in the reply content being not appropriate or failing to meet user needs, making it difficult to accurately identify the user's intention, resulting in reduced accuracy of the final recognition. Summary of the invention

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for optimizing intelligent customer service dialogue strategies based on deep learning, including: S1, obtaining the first text information of the previous round of user text, and obtaining a text reply library corresponding to the previous round of user text based on the first text information.

[0006] S2, according to the text reply library corresponding to each round of user text, sentiment matching is performed on the identified keywords to determine the sentiment tendency and intention of the user text.

[0007] S3, based on the sentiment tendency and intention of the user text, generates the customer service reply text and determines the predicted sentiment features and predicted reply features of the reply text.

[0008] S4, obtaining the user's feedback information, matching the user's feedback information with the customer service's reply text, and obtaining a matched target sequence; according to the matched target sequence, obtaining a customer service dialogue strategy.

[0009] The beneficial effects of the present invention are as follows: the present invention obtains the user text input by the user in the previous round, generates a text reply library corresponding to the user text, and identifies the text reply library input in each round according to the difference in the text input by the user in each round, so as to determine the emotional tendency and intention of the current user text; according to the emotional tendency and intention of the user text, the customer service reply text is adjusted to obtain the reply text that is most relevant to the user text at present, and after matching the reply text according to the feedback information, the most appropriate customer service dialogue strategy is found, thereby improving the accuracy and efficiency of the customer service dialogue. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0011] Figure 1 It is a flowchart of an intelligent customer service dialogue strategy optimization method based on deep learning.

[0012] Figure 2 The present invention is a flowchart of step S1 of an intelligent customer service dialogue strategy optimization method based on deep learning.

[0013] Figure 3 The present invention is a flowchart of step S3 of an intelligent customer service dialogue strategy optimization method based on deep learning.

[0014] Figure 4 The present invention is a flowchart of step S4 of an intelligent customer service dialogue strategy optimization method based on deep learning. DETAILED DESCRIPTION

[0015] The embodiments of the present invention are described in detail below. The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. If no specific techniques or conditions are specified in the embodiments, the techniques or conditions described in the literature in the art or the product specifications are used.

[0016] See also Figure 1 , a method for optimizing intelligent customer service dialogue strategy based on deep learning, comprising: S1, obtaining the first text information of the previous round of user text, and obtaining a text reply library corresponding to the previous round of user text according to the first text information.

[0017] S2, according to the text reply library corresponding to each round of user text, sentiment matching is performed on the identified keywords to determine the sentiment tendency and intention of the user text.

[0018] S3, based on the sentiment tendency and intention of the user text, generates the customer service reply text and determines the predicted sentiment features and predicted reply features of the reply text.

[0019] S4, obtaining the user's feedback information, matching the user's feedback information with the customer service's reply text, and obtaining a matched target sequence; according to the matched target sequence, obtaining a customer service dialogue strategy.

[0020] like Figure 2 As shown, the processing method of obtaining the text reply library corresponding to the previous round of user text in step S1 is: S11, obtaining the first text information input by the user in the previous round of conversation, the first text information includes all the text input by the user in the previous round, and the first text information also includes the labels and names of the products and stores currently input by the user when the user is inputting; at this time, obtaining the user's previous round of user text means the content input by the user in the most recent interaction with customer service.

[0021] S12, obtaining keywords in the first text information, and determining the occurrence probability and relative weight of the keywords in the first text information in the previous round of user text.

[0022] S13, based on the obtained keywords, obtain reply content related to the keywords of the first text message from the preset reply library, sort the reply content according to the appearance probability and relative weight of the keywords of the first text message, and use the sorted reply content as the text reply library corresponding to the previous round of user text.

[0023] At this time, after obtaining the first text information based on the text input by the user, the total number of times each keyword appeared in the user's last round and the relative weight set for each keyword in the preset scenario are calculated in turn. The relative weight is pre-set in the database. At this time, the relative weight is obtained by comparing the keywords that appear, and according to the size of the relative weight, the keywords are used to obtain the reply content from the database, and the obtained text is sorted according to the probability of occurrence and relative weight of the keywords to obtain the text reply library that currently needs to be processed.

[0024] The method for obtaining keywords in the first text information includes: obtaining labels and names of products related to the user input in the first text information, determining first related words of the labels and names of the products, and obtaining second related words corresponding to all texts input by the user in the first text information. The first related words are extracted from the labels and names of the products related to the user input, and the second related words are obtained from the text input by the user, and the words that the user focuses on when inputting are determined, and there are multiple groups of first related words and second related words.

[0025] Compare multiple groups of first related words and second related words to determine the similarity between the first related words and the second related words, and output a group of first related words and second related words whose similarity between the first related words and the second related words is greater than a preset similarity as keywords of the first text information. The preset similarity is set to 0.6, which is used to screen out more relevant vocabulary parts from the first related words and the second related words.

[0026] The vectors of the first related word and the second related word are converted into word vectors to calculate the similarity, and the calculated similarity is cosine similarity; the first related word and the second related word are both extracted from the first text information to obtain corresponding data, and this part of the data is compared with the preset text to obtain.

[0027] When using the keywords in the first text information to obtain the user's text reply library, the processing method of the text reply library also includes: identifying the scenario factors, state transition factors and execution actions corresponding to the previous round of user text based on the acquired text reply library, determining the state transition probability corresponding to the execution action under each scenario factor and the observation results after the state transition factor changes, and obtaining the execution vector value of each execution action based on the state transition probability and the observation results.

[0028] The state transition probability is expressed as: P(s t+1 |c t ,a t )=Pr(S t+1 =s t+1 |C t =c t ,A t =a t ), where P(s t+1 |c t ,a t ) is the state transition probability, s t+1 Indicates the state transition factor at the next moment, c t is the scene factor at the current moment, a t is the execution action taken at the current moment, P(s t+1 |c t ,a t ) represents the current execution scenario factor c t and perform action a t The state transfer factor s is transferred to the next moment t+1 The probability, Pr(S t+1 =s t+1 |C t =c t ,A t =a t ) is represented as the scene factor C t=c t and the execution action is A t =a t Under the condition of t+1 For t+1 The probability of , .

[0029] The observation result is expressed as, according to the state transition factor and the observation result in the next moment, the observation probability corresponding to the observation result is determined: P(O t+1 |s t+1 )=Pr(O t+1 =o t+1 |S t+1 =s t+1 ); where P(O t+1 |s t+1 ) represents the observation probability, s t+1 Indicates the state transition factor at the next moment, o t+1 represents the observation result at the next moment, which is obtained by observing whether the user is satisfied after the state transition factorization and whether an error is prompted at present; P(O t+1 |s t+1 ) represents the state transition factor s at the next moment t+1 Observed o t+1 The probability of the corresponding observation, Pr(O t+1 =o t+1 |S t+1 =s t+1 ) indicates that the state transition factor is S t+1 =s t+1 Under the condition of t+1 for o t+1 probability.

[0030] The execution vector value is expressed as, based on the observation probability and state transition probability corresponding to the observation result, the execution vector value is obtained.

[0031] Among them, Q(c t ,a t ) represents the execution vector value, i.e., the scenario factor c t and perform action a t The expected return of R t+1 represents the immediate reward, which is a scalar value representing the reward immediately after taking the action; γ is the discount factor, which is a scalar value between 0 and 1 that measures the importance of future rewards; s t+1,k represents the kth state transition factor at the next moment, k = 1, 2,,, K, K represents the number of state transition factors existing at the next moment; c tis the scene factor at the current moment, a t is the execution action taken at the current moment; o t+1,g represents the g-th observation result at the next moment, g = 1, 2,,, G, G represents the number of observation results at the next moment; a′ represents the number of observation results that can be given to s t+1,k The execution action that brings the maximum gain value; max a′ Q(s t+1,k ,a′) indicates that the state transition factor is s t+1,k and the execution action is a′, Q(s t+1,k ,a′) corresponds to the maximum execution vector value; P(s t+1,k |c t ,a t ) represents the current execution scenario factor c t and perform action a t The kth state transfer factor s is transferred to the next moment t+1,k The probability, P(O t+1,g |s t+1,k ) represents the kth state transition factor s at the next moment t+1,k Observed o t+1,g The probability of the corresponding g-th observation.

[0032] According to the obtained execution vector value, the execution action corresponding to the maximum execution vector value is selected, and the reply text library containing the execution action is used as the reply text library finally outputted in this step.

[0033] At this time, the execution vector value corresponding to the reply text library is calculated to determine the main content of the current user's reply in the previous round, and the text that the user needs to reply is obtained based on this main content, so that a reply can be made to the user's specific text.

[0034] Scenario factors represent conversation scenarios, user context, and time factors.

[0035] Dialogue scenario: a specific scenario in which users interact with customer service, such as shopping consultation, after-sales service, technical support, etc.

[0036] User background: User’s personal information, historical interaction records, purchase records, etc. This background information will affect the content and method of the conversation.

[0037] Time factor: The time of day, night, weekdays, weekends, etc. when the conversation takes place may affect the user's patience and expected response speed.

[0038] The state transition factor represents the user state change, customer service state change and conversation state change.

[0039] User state changes: During a conversation, a user may change from confused to satisfied, or from satisfied to dissatisfied.

[0040] Customer service status change: Customer service may go from not understanding the user's problem at the beginning to gradually understanding it and providing solutions.

[0041] Dialogue state changes: A conversation may progress from an initial inquiry phase to a problem-solving phase and then to a conclusion phase.

[0042] Execution actions represent user actions and customer service actions.

[0043] User actions: such as asking questions, providing information, confirming, denying, etc.

[0044] Customer service actions: such as answering questions, providing suggestions, requesting more information, confirming user intent, etc.

[0045] The observation probability represents the difference in user feedback, system feedback, and conversation content due to changes in state transition factors.

[0046] User feedback: The user's tone, words, emoticons, etc. can reflect their satisfaction and emotional state.

[0047] System feedback: such as system response time, whether error prompts appear, etc.

[0048] Conversation content: The content of the conversation can be used to observe whether the user’s needs are met and whether the problem is solved.

[0049] Based on the above analysis, we can know the user's preference for the current reply content under the corresponding action, so as to improve the accuracy of the reply to the user based on these identified preferences.

[0050] The method for acquiring the emotional tendency in step S2 includes: acquiring the second text information based on the text reply library of each round, the second text information being the emotional words existing in each reply of the user in the text reply library, and being used to determine whether the user's emotions and intentions change when communicating in reply.

[0051] The acquired second text information is sorted according to the round and time of user replies, the sentiment score of each paragraph when the user replies is calculated in turn, and a sentiment trend curve is formed according to the sentiment score of each paragraph and the time of the reply.

[0052] The slope value of each corresponding time point in the sentiment trend curve is calculated, and the standard deviation and variance of the slope value are obtained. Based on the obtained standard deviation and variance, it is determined whether the user's intention and sentiment have changed, and the sentiment tendency of the user's text is calculated.

[0053] The user's emotional tendency is determined by the standard deviation and variance of the slope value. If the slope value shows an overall upward trend and the standard deviation and variance of the slope value are greater than or equal to the preset score, the user's emotional tendency can be considered to be positive. If the slope value shows an overall upward trend and the standard deviation and variance of the slope value are less than the preset score, the current user's emotional tendency is considered to be neutral. If the slope value shows an overall downward trend and the standard deviation and variance of the slope value are less than the preset score, the user's emotional tendency can be considered to be negative.

[0054] The sentiment score of the second text information is obtained by comparing the second text information with a set sentiment dictionary, identifying sentiment words in the second text information, and obtaining the scores of the sentiment words from the set sentiment dictionary. The sentiment score of each paragraph is the sum of the scores of all sentiment words in the paragraph.

[0055] Determine the slope value of each corresponding time point in the emotional trend curve after each reply of the user, and identify the change of the slope value at each time point. If the currently acquired emotional trend curve has an extreme point, obtain the adjacent points of the extreme point, and determine the change in the emotional tendency of the current user according to the difference between the slope values ​​of the adjacent points and the extreme point, and use the tendency identified at this time as the emotional tendency of the user text; if the currently acquired emotional trend curve does not have an extreme point, obtain the average value of the slope value in the emotional trend curve as the emotional tendency of the current user text.

[0056] The method of obtaining the intention of the user text is as follows.

[0057] Perform semantic analysis on the second text information to determine the intent region in the second text information corresponding to the target text, where the target text represents the vocabulary selected from the second text information for semantic analysis, and the intent region is used to determine the scope of identifying user intent in the second text information; extract entity words corresponding to the target text from the intent region, and use the extracted entity words and the identifier of the intent region as the first analysis result of the semantic analysis.

[0058] The intention area is implemented by extracting judgment words related to the target text from the second text information. The judgment words can represent the tendency of the current intention and the specific content of the intention, and the paragraph area containing the judgment words is used as the intention area corresponding to the target text.

[0059] The entity words in the intent area are used to represent the main entities that the user wants to express in the intent area. For example, when making product inquiries, the entity words extracted are the current product, product properties, labels, and other words that directly represent the product. The identifier of the intent area indicates what category the current identification intent area belongs to, such as consulting questions, feedback on product quality, querying product information, etc., which is used to classify the text currently entered by the user.

[0060] If there is no entity word in the intent region, the entity word of the adjacent intent region and the identifier of the intent region are used as the second analysis result of the semantic analysis. If the corresponding entity word cannot be identified at this time, the entity word in the text closest to the current intent region is selected as the output; the content after comparing the first analysis result and the second analysis result is used as the intent of the user text.

[0061] At this time, the method of identifying the user is to identify whether there are keywords that can express the user's intention in the second text information, find the user's intention based on the judgment words that appear in this part, and understand the interpretation of the judgment words as the current user's intention.

[0062] like Figure 3 As shown, in step S3, the method of generating the customer service reply text based on the sentiment tendency and intention of the user text is as follows.

[0063] S31, convert the sentiment tendency and intention of the user text into a word vector, and obtain the word vector corresponding to the preset reply template.

[0064] S32, calculate the distance and completeness between the word vector of the user text and the word vector of the reply template, and determine the comprehensive score of the distance and completeness, and use the reply template corresponding to the group of word vectors with the largest comprehensive score of each distance and completeness as the customer service reply text.

[0065] The distance between word vectors is expressed as the average cosine similarity between the word vector corresponding to the user text and the word vector corresponding to the reply template.

[0066] The completeness between word vectors is expressed as the matching ratio between the word vector corresponding to the user text and the word vector corresponding to the reply template.

[0067] Given the word vector u corresponding to the user text, 1 ,u 2 ,,,u n ], where n is the number of words in the user text; the word vector corresponding to the preset reply template is represented by r = [r 1 ,r 2 ,,,r m ], where m is the number of words in the response template.

[0068] The distance is expressed as: Among them, D(u,r) represents the distance between word vectors, u i represents the i-th word vector in the user text, r j represents the jth word vector in the reply template, n is the number of words in the user text, m is the number of words in the reply template, i = 1, 2,,,n, j = 1, 2,,,m; cos(u i ,rj ) represents the word vector u i and word vector r j The cosine similarity between 1≤j≤m Indicates finding the word vectors of all reply templates that match u i The most similar word vector.

[0069] The completeness is expressed as: Among them, C(u,r) represents the completeness between word vectors; 1 {·} It represents the indicator function, which returns 1 if the condition in the brackets is met, otherwise it returns 0. θ represents the pre-set threshold value, which is used to determine whether the similarity between two vectors is high enough.

[0070] Finally, the weighted sum of distance and completeness is obtained to get the comprehensive score related to the word vector. According to the obtained comprehensive score, the reply template with the largest comprehensive score value is selected as the customer service reply text. That is, the comprehensive score of distance and completeness is expressed as shown in the following formula.

[0071] S(u,r)=ω 1 ×D(u,r)+ω 2 ×C(u,r); where S(u,r) represents the comprehensive score of the word vector, D(u,r) represents the distance between word vectors, C(u,r) represents the completeness between word vectors, ω 1 Represents the weight of the distance between word vectors, ω 2 Represents the weight of the completeness between word vectors.

[0072] By converting the sentiment and intention of the user text into word vectors and matching them with the word vectors corresponding to the preset reply templates, a more personalized reply text that meets user needs can be generated; the distance takes into account the similarity between the user text and the reply template, while the completeness takes into account the coverage of the vocabulary in the user text in the reply template. The combination of the two can more comprehensively evaluate the quality of the reply template; when a new reply template needs to be added, it only needs to be converted into a word vector and the distance and completeness between the word vector and the user text are calculated, without the need for a large number of modifications to the entire system, so that a reply text that meets the current user can be obtained more quickly.

[0073] The implementation method of determining the predicted sentiment features and predicted reply features of the reply text is as follows.

[0074] The predicted emotional features represent the emotional features and emotional tendencies that may appear after the user obtains the reply text; the predicted reply features represent the reply features that may appear after the user receives the reply text.

[0075] The method for obtaining the predicted emotional features is to use the word vectors corresponding to the emotional tendency of the user text in the reply text as the basis for predicting the emotional features, calculate the scores of the predicted emotional features, and use the calculated scores of the predicted emotional features as the predicted emotional features of the reply text at this time.

[0076] For the predicted reply feature, the content in the reply text corresponding to the user's previous input text is obtained as the predicted reply feature, that is, the predicted reply feature represents the words most relevant to the reply text and the user's input text.

[0077] Compare the currently recognized predicted emotion features and predicted reply features with the actual user input content. If there is a mismatch, re-judge the extracted features to ensure that the extracted predicted emotion features and predicted reply features are closely related to the user, and set weights for the extracted features based on the context and time information of the user text. According to the weighted sum of all extracted features, obtain the features that best match the weighted sum to meet the needs of user input.

[0078] like Figure 4 As shown, the implementation method of matching the user's feedback information with the customer service's reply text in step S4 includes: S41, performing feature processing on the feedback information according to a feature extraction algorithm to determine feedback features corresponding to the user's feedback information.

[0079] S42, clustering the feedback features, the predicted sentiment features of the reply text, and the predicted reply features into a plurality of question clusters.

[0080] S43: Determine a sequence identifier corresponding to the feedback feature according to the user's feedback information.

[0081] S44, processing the problem clusters in sequence according to the sequence identifiers to obtain a target sequence.

[0082] When the feedback information is processed with features, the feedback information is mapped with preset words, the semantic relationship between the input feedback information and the set words is clarified, and this part of the semantic relationship is used as the feedback feature.

[0083] The method for processing the problem clustering cluster is as follows: count the number of times the combinations in the problem clustering cluster appear together, and determine the order of sequential identifiers when the combinations in the problem clustering cluster appear together; according to the order of sequential identifiers, determine the combinations in the problem clustering cluster as related clustering sets; determine the index value of the related clustering set in the corresponding sequential identifier; and take the related clustering set whose index value is greater than the preset index value as the target sequence.

[0084] The combination in the question clustering cluster is represented as a set of features after clustering feedback features, predicted sentiment features and predicted reply features. The clustering method at this time is to group the feedback features, predicted sentiment features and predicted reply features according to the same mapped semantic relationship, so as to represent the relationship between the user's feedback information and the current generated text.

[0085] The above sequence identifiers include: emotional tendency, problem solving degree, response speed, communication clarity and service attitude. The sequence identifiers for emotional tendency, problem solving degree, response speed, communication clarity and service attitude are set as 1, 2, 3, 4, 5, respectively; to determine the corresponding problem in each feedback information.

[0086] The user's emotional tendency is the primary feature of feedback information, which reflects the user's overall feeling towards the customer service conversation, such as satisfaction, dissatisfaction, neutrality, etc.; problem resolution is a key indicator to measure whether the customer service conversation has effectively solved the user's problem, which reflects the customer service's professional ability and problem-solving ability; response speed refers to the speed at which customer service responds to user problems. A quick response can improve user satisfaction and trust; communication clarity refers to the clarity of information expressed by customer service in the conversation. Clear communication can reduce misunderstandings and ambiguity and improve conversation efficiency; service attitude refers to the attitude and behavior of customer service in the conversation. A good service attitude can improve user satisfaction and loyalty.

[0087] Through the above sequence identification and explanation, we can better understand and utilize user feedback information to evaluate and optimize customer service dialogue strategies.

[0088] Sentiment tendency: Calculate the average sentiment tendency of the feedback in each cluster, use the sentiment analysis tool to calculate the sentiment score of each feedback, and then take the average.

[0089] Problem solving degree: Analyze the feedback in each cluster to determine whether the problem has been solved. If most of the feedback indicates that the problem has been solved, the problem solving degree is high.

[0090] Response speed: Calculate the response time of feedback in each cluster and calculate the average time from the customer initiating a request to receiving a reply.

[0091] Communication clarity: Analyze the clarity of responses within each cluster to assess whether there are clear answers or instructions in the responses.

[0092] Service attitude: Evaluate the attitude of the customer service staff in each cluster and judge whether the customer service staff is friendly and patient based on the wording in the feedback.

[0093] Assume that there are the following clusters (Cluster1, Cluster2, Cluster3):

[0094] Cluster: Feedback feature: Customers wait for a long time, Prediction emotion feature: Negative emotion, Prediction reply feature: Explain the reason and provide compensation plan.

[0095] Cluster 2: Feedback feature: Customers are dissatisfied with the product’s features, Prediction sentiment feature: Neutral, Prediction response feature: Provide technical solutions.

[0096] Cluster 3: Feedback feature: Customers are dissatisfied with the attitude of service personnel, predicted emotional feature: negative emotions, predicted response feature: Apology and promise to improve.

[0097] Sentiment tendency: Cluster 1: the average sentiment tendency score is low (e.g. 2.5 / 5), Cluster 2: the average sentiment tendency score is medium (e.g. 3.5 / 5), Cluster 3: the average sentiment tendency score is low (e.g. 2.5 / 5).

[0098] Problem Solving Degree: Cluster 1: low (e.g., 2 / 5), because the feedback indicated that the waiting time was long but the problem was not necessarily solved, Cluster 2: high (e.g., 4 / 5), because a technical solution was provided, Cluster 3: medium (e.g., 3 / 5), because although an apology was given, the fundamental problem was not solved.

[0099] Response speed: Cluster 1: Slow (e.g. 2 / 5) because customers had a long waiting time, Cluster 2: Medium (e.g. 3 / 5) because the response speed was not specifically mentioned, Cluster 3: Slow (e.g. 2 / 5) because customers were dissatisfied with the waiting time.

[0100] Communication clarity: Cluster 1: medium (e.g., 3 / 5) because a compensation plan was provided, Cluster 2: high (e.g., 4 / 5) because a specific solution was provided, Cluster 3: medium (e.g., 3 / 5) because although an apology was given, there was a lack of a specific action plan.

[0101] Service attitude: Cluster 1: medium (e.g. 3 / 5), because although the waiting time was long, compensation was provided, Cluster 2: high (e.g. 4 / 5), because professional solutions were provided, Cluster 3: low (e.g. 2 / 5), because customers were dissatisfied with the attitude of the service staff.

[0102] In this way, five sequentially identified index values ​​can be calculated for each cluster. These index values ​​can help us understand the performance of different clusters in these aspects and take corresponding improvement measures accordingly.

[0103] In step S4, the customer service dialogue strategy is obtained according to the matched target sequence as follows:

[0104] Obtain the index value of the sequence mark corresponding to the target sequence to obtain the comprehensive index value. The index value of the sequence mark includes the index values ​​corresponding to emotional tendency, problem solving degree, response speed, communication clarity and service attitude. According to the obtained comprehensive index value, select the corresponding customer service dialogue strategy from the preset dialogue strategy scheme.

[0105] Com=EO×α 1 +PR×α 2 +RS×α 3 +CC×α 4 +SA×α 5 ; Among them, Com is the comprehensive index value, EO is the index value corresponding to the emotional tendency, α 1 is the weight corresponding to the sentiment tendency, PR is the index value corresponding to the problem solving degree, α 2 is the weight corresponding to the problem solving degree, RS is the index value corresponding to the response speed, α 3 is the weight corresponding to the response speed, CC is the index value corresponding to the communication clarity, α 4 is the weight corresponding to communication clarity, SA is the index value corresponding to service attitude, α 5 is the weight corresponding to service attitude.

[0106] The corresponding customer service dialogue strategy is selected from the preset dialogue strategy schemes by calculating the index value in the current sequence identifier and comparing it with the preset dialogue strategy schemes in the order of the sequence identifiers. When the confidences of all sequence identifiers and preset dialogue strategy schemes are greater than the preset confidences, the confidences of the sequence identifiers and preset dialogue strategy schemes corresponding to the comprehensive index value are calculated, and the dialogue strategy scheme that is greater than the preset comprehensive confidence at this time is used as the currently used customer service dialogue strategy. The preset confidence value can be 0.6.

[0107] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention and they are still covered by the protection scope of the present invention.

Claims

1. A method for optimizing intelligent customer service dialogue strategy based on deep learning, characterized in that: include: S1, obtaining the first text information of the previous round of user text, and obtaining a text reply library corresponding to the previous round of user text according to the first text information; S2, according to the text reply library corresponding to each round of user text, sentiment matching is performed on the identified keywords to determine the sentiment tendency and intention of the user text; S3, based on the sentiment tendency and intention of the user text, generates the customer service reply text and determines the predicted sentiment features and predicted reply features of the reply text; S4, obtaining the user's feedback information, matching the user's feedback information with the customer service's reply text, and obtaining a matched target sequence; according to the matched target sequence, obtaining a customer service dialogue strategy.

2. According to claim 1, a method for optimizing intelligent customer service dialogue strategy based on deep learning is characterized in that: The processing method of obtaining the text reply library corresponding to the previous round of user text in step S1 is: S11, obtaining first text information input by the user in the last round of conversation, where the first text information includes all texts input by the user in the last round; S12, obtaining keywords in the first text information, and determining the occurrence probability and relative weight of the keywords in the first text information in the previous round of user texts; S13, based on the obtained keywords, obtain reply content related to the keywords of the first text message from the preset reply library, sort the reply content according to the appearance probability and relative weight of the keywords of the first text message, and use the sorted reply content as the text reply library corresponding to the previous round of user text.

3. The method for optimizing intelligent customer service dialogue strategy based on deep learning according to claim 2, characterized in that: Methods for obtaining the keywords in the first text information include: Obtaining the label and name of the product related to the user input in the first text information, determining the first related words of the product label and name, and obtaining the second related words corresponding to all the texts input by the user in the first text information, wherein there are multiple groups of the first related words and the second related words; Compare multiple groups of first related words and second related words to determine the similarity between the first related words and the second related words, and output a group of first related words and second related words whose similarity between the first related words and the second related words is greater than a preset similarity as keywords of the first text information.

4. According to the deep learning-based intelligent customer service dialogue strategy optimization method of claim 1, it is characterized in that: The method of obtaining the sentiment tendency in step S2 includes: According to each round of text reply library, second text information is obtained, where the second text information is the sentiment words present in each reply of the user in the text reply library; The acquired second text information is sorted according to the round and time of the user's reply, the sentiment score of each paragraph when the user makes a reply is calculated in sequence, and a sentiment trend curve is formed according to the sentiment score of each paragraph and the time of the reply; The slope value of each corresponding time point in the sentiment trend curve is calculated, and the standard deviation and variance of the slope value are obtained. Based on the obtained standard deviation and variance, the sentiment tendency of the user text is calculated.

5. The method for optimizing intelligent customer service dialogue strategy based on deep learning according to claim 4, characterized in that: The method for obtaining the user's text intent is: Performing semantic analysis on the second text information, determining an intention region in the second text information corresponding to a target text, wherein the target text represents words selected from the second text information for semantic analysis, extracting entity words corresponding to the target text from the intention region, and using the extracted entity words and identifiers of the intention region as first analysis results of the semantic analysis; If there is no entity word in the intent region, the entity word of the adjacent intent region and the identifier of the intent region are used as the second analysis result of the semantic analysis; The content after comparing the first analysis result and the second analysis result is used as the intention of the user text.

6. The method for optimizing intelligent customer service dialogue strategy based on deep learning according to claim 1, characterized in that: In step S3, based on the sentiment and intention of the user text, the customer service reply text is generated in the following manner: S31, converting the sentiment tendency and intention of the user text into a word vector, and obtaining the word vector corresponding to the preset reply template; S32, calculate the distance and completeness between the word vector of the user text and the word vector of the reply template, and determine the comprehensive score of the distance and completeness, and use the reply template corresponding to the group of word vectors with the largest comprehensive score of each distance and completeness as the customer service reply text.

7. The method for optimizing intelligent customer service dialogue strategy based on deep learning according to claim 6, characterized in that: Said The distance is expressed as: Among them, D(u,r) represents the distance between word vectors, u i represents the i-th word vector in the user text, r j represents the jth word vector in the reply template, n is the number of words in the user text, m is the number of words in the reply template, i = 1, 2,,,n, j = 1, 2,,,m; cos(u i ,r j ) represents the word vector u i and word vector r j The cosine similarity between ; The completeness is expressed as: Among them, C(u,r) represents the completeness between word vectors; 1 {·} represents the indicator function, and θ represents the preset threshold.

8. The method for optimizing intelligent customer service dialogue strategy based on deep learning according to claim 7, characterized in that: Obtain the weighted sum of distance and completeness to obtain a comprehensive score of distance and completeness; S(u,r)=ω1×D(u,r)+ω2×C(u,r); Among them, S(u,r) represents the comprehensive score of the word vector, D(u,r) represents the distance between word vectors, C(u,r) represents the completeness between word vectors, ω1 represents the weight of the distance between word vectors, and ω2 represents the weight of the completeness between word vectors.

9. The method for optimizing intelligent customer service dialogue strategy based on deep learning according to claim 1, characterized in that: The implementation method of matching the user's feedback information with the customer service's reply text in step S4 includes: S41, performing feature processing on the feedback information according to a feature extraction algorithm to determine feedback features corresponding to the user's feedback information; S42, clustering the feedback features, the predicted sentiment features of the reply text, and the predicted reply features into a plurality of question clusters; S43, determining a sequence identifier corresponding to the feedback feature according to the user's feedback information; S44, processing the problem clusters in sequence according to the sequence identifiers to obtain a target sequence.

10. The method for optimizing intelligent customer service dialogue strategy based on deep learning according to claim 9, characterized in that: In step S4, the customer service dialogue strategy is obtained according to the matched target sequence as follows: Obtain the index value of the sequence mark corresponding to the target sequence to obtain the comprehensive index value. The index value of the sequence mark includes the index values ​​corresponding to the emotional tendency, problem solving degree, response speed, communication clarity and service attitude. According to the obtained comprehensive index value, select the corresponding customer service dialogue strategy from the preset dialogue strategy scheme; Com=EO×α1+PR×α2+RS×α3+CC×α4+SA×α5; Among them, Com is the comprehensive index value, EO is the index value corresponding to emotional tendency, α1 is the weight corresponding to emotional tendency, PR is the index value corresponding to problem solving degree, α2 is the weight corresponding to problem solving degree, RS is the index value corresponding to response speed, α3 is the weight corresponding to response speed, CC is the index value corresponding to communication clarity, α4 is the weight corresponding to communication clarity, SA is the index value corresponding to service attitude, and α5 is the weight corresponding to service attitude.

Images