Intelligent customer service question and answer optimization method and system based on sentiment analysis

Abstract

The invention discloses an intelligent customer service question and answer optimization method and system based on sentiment analysis, and relates to the technical field of language data processing, and the specific steps are as follows: reading voice data, and carrying out sentiment analysis and feature extraction on the voice data; judging the priority level of the voice data based on the extracted features; performing demand prediction on the voice data in sequence based on the priority level; and generating personalized reply content according to the extracted features, the prediction demand and the historical interaction record. By introducing emotion recognition, the emotion change of the user can be accurately captured, the reply tone and content are adjusted accordingly, more intimate service is provided, the negative emotion of the user can be relieved, and the trust feeling and satisfaction degree of the user to a power enterprise are increased.

Description

Technical Field

[0001] The present invention relates to the technical field of language data processing, and more specifically to an intelligent customer service question and answer optimization method and system based on sentiment analysis. Background Art

[0002] With the rapid development of information technology, the power industry is also gradually transforming towards intelligence and informatization. As an important bridge between power companies and users, the quality of customer service directly affects the image of the company and user satisfaction. Traditional customer service usually relies on manual customer service, but with the growth of the user base and the diversification of service needs, it is difficult to meet efficient and timely service requirements with manpower alone. Therefore, intelligent customer service systems came into being and have been widely used in the power industry.

[0003] Existing intelligent customer service systems mainly rely on preset question libraries and keyword matching to answer user questions. Although this method can quickly respond to users' common questions, it has the following shortcomings:

[0004] Lack of emotion recognition: Existing systems can usually only recognize users' text messages, but ignore the emotional color. For example, when users express dissatisfaction or urgent needs, the system cannot accurately capture and respond to these emotions, causing users to feel that the service is indifferent and uncaring;

[0005] Templated answers: Due to the lack of understanding of the user's specific context, intelligent customer service often provides standardized answers, which are difficult to meet the user's personalized needs. This templated answer method makes users feel like they are communicating with a machine and lacks a humanized service experience;

[0006] Inflexible interaction: The existing system is rigid when dealing with complex questions or multi-round conversations. Once the user's questions deviate from the preset question path, the system will find it difficult to respond effectively, resulting in a poor user experience.

[0007] In view of the above problems, how to develop a power grid intelligent customer service system that can understand and respond to user emotions and provide more personalized and flexible services has become an urgent problem that technicians in this field need to solve. Summary of the invention

[0008] In view of this, the present invention provides an intelligent customer service question and answer optimization method and system based on sentiment analysis, which overcomes the above-mentioned defects.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] An intelligent customer service question and answer optimization method based on sentiment analysis, the specific steps are:

[0011] Reading voice data, and performing sentiment analysis and feature extraction on the voice data;

[0012] Determining the priority level of the voice data based on the extracted features;

[0013] Performing demand forecasting on the voice data in order based on the priority levels;

[0014] Generate personalized reply content based on the extracted features, predicted needs and historical interaction records.

[0015] Optionally, the specific steps of sentiment analysis and feature extraction are:

[0016] Recognize the read voice data to generate text data;

[0017] Preprocessing the speech data and the text data to obtain speech preprocessing data and text preprocessing data;

[0018] Extracting features from the speech preprocessing data and the text preprocessing data, wherein the extracted features include displayed emotion features, implicit emotion features, and demand targets; wherein the displayed emotion features are displayed emotion text features, and the implicit emotion features include implicit emotion speech features and implicit emotion text features;

[0019] The displayed emotion feature and the implicit emotion feature are fused to obtain the emotion state.

[0020] Optionally, the preprocessing of the speech data includes speech denoising, silence detection, speech sentence segmentation and normalization.

[0021] Optionally, the preprocessing of the text data includes text denoising, text segmentation, part-of-speech tagging, named entity recognition and syntactic analysis.

[0022] Optionally, the step of acquiring the emotional state is:

[0023] The multi-head mutual attention mechanism is used to learn the interactive information between implicit emotional speech features and explicit emotional text features, and the interactive information between implicit emotional speech features and implicit emotional text features;

[0024] The cross-attention mechanism is used to learn the attention weights between the interaction information and the explicit emotional text features to identify the emotional state.

[0025] Optionally, the specific steps of determining the priority level are:

[0026] Quantify the demand target according to a preset urgency quantification table to obtain a demand quantification value;

[0027] Quantifying the emotional state according to a preset emotional quantification table to obtain an emotional quantification value;

[0028] The priority level is obtained by using a priority rule based on the demand quantification value and the emotion quantification value.

[0029] Optionally, the specific steps of demand forecasting are:

[0030] Read the demand target, the emotional state, and the historical interaction record of the voice data in order based on the priority level;

[0031] A demand prediction result is generated using a demand prediction model according to the demand target, the emotional state and the historical interaction record.

[0032] Optionally, the steps for obtaining personalized reply content are:

[0033] Calling multiple pending response plans according to the demand forecast results;

[0034] Invoking a response template according to the emotional state;

[0035] The solution to be responded is selected using the demand target, and a reply voice is generated from the solution to be responded according to the template to be responded.

[0036] An intelligent customer service question and answer optimization system based on sentiment analysis, comprising:

[0037] A feature extraction module is used to read voice data and perform sentiment analysis and feature extraction on the voice data;

[0038] A priority level determination module, used for determining the priority level of the speech data according to the extracted features;

[0039] A demand prediction module, used for performing demand prediction on the voice data in order based on the priority levels;

[0040] The reply module is used to generate personalized reply content based on the extracted features, predicted needs and historical interaction records.

[0041] Optionally, a system optimization module is also included, specifically: generating a service evaluation report according to the collected feedback data and sentiment change trends, and optimizing the system based on the service evaluation report.

[0042] It can be seen from the above technical solutions that the present invention provides an intelligent customer service question and answer optimization method and system based on sentiment analysis, which has the following beneficial effects compared with the prior art:

[0043] 1. By introducing emotion recognition, we can accurately capture the user's emotional changes and adjust the tone and content of the reply accordingly to provide more considerate services, which helps to alleviate the user's negative emotions and increase their trust and satisfaction with the power company.

[0044] 2. Use natural language processing technology to optimize the multi-round dialogue management mechanism, so that the machine can more flexibly respond to continuous inquiries or emergencies in complex scenarios, increase the dialogue process, reduce the workload of manual customers, improve response efficiency, and greatly improve the overall effectiveness of customer service.

[0045] 3. Continuously optimize system performance through feedback data, which is conducive to quickly adapting to market changes and technological advances. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0047] Figure 1 A schematic diagram of the method provided by the present invention;

[0048] Figure 2 This is a schematic diagram of the system structure provided by the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] On the one hand, this embodiment discloses an intelligent customer service question and answer optimization method based on sentiment analysis, such as Figure 1 As shown, the specific steps are:

[0051] Step 1: Read the voice data and perform sentiment analysis and feature extraction on the voice data;

[0052] Step 2: Determine the priority of the voice data based on the extracted features;

[0053] Step 3: Predict the demand for voice data in order based on the priority level;

[0054] Step 4: Generate personalized response content based on extracted features, predicted needs, and historical interaction records.

[0055] In one embodiment, the specific steps of sentiment analysis and feature extraction are:

[0056] Step 11: Recognize the read voice data to generate text data;

[0057] Step 12: pre-process the voice data and text data to obtain voice pre-processed data and text pre-processed data;

[0058] Step 13: extracting features from the speech preprocessed data and the text preprocessed data, wherein the extracted features include explicit emotion features, implicit emotion features, and demand targets;

[0059] Step 14: Fuse the explicit emotion features and implicit emotion features to obtain the emotion state.

[0060] Furthermore, in step 11, the read voice data is converted into text data using automatic speech recognition (ASR) technology.

[0061] In one embodiment, the preprocessing of the speech data in step 12 includes speech denoising, silence detection, speech sentence segmentation and normalization.

[0062] Furthermore, the preprocessing of the speech data is specifically as follows:

[0063] Speech denoising: denoising speech data based on bilinear transformation method;

[0064] Silence detection: detect whether there are silence segments in the voice data, and if so, remove them to improve the efficiency of voice recognition and processing;

[0065] Speech Sentence Segmentation: Automatically segment speech data into meaningful sentence units, using methods such as endpoint detection and speech activity detection;

[0066] Normalization: Adjust the signal amplitude of speech data to a certain range to eliminate volume differences between different speakers.

[0067] In one embodiment, the preprocessing of the text data in step 12 includes text denoising, text segmentation, part-of-speech tagging, named entity recognition and syntactic analysis.

[0068] Furthermore, the preprocessing of text data is specifically as follows:

[0069] Text denoising: remove non-text elements (such as HTML tags, special symbols, extra spaces, etc.) from text data, correct spelling errors, unify capitalization and punctuation, etc.;

[0070] Text segmentation: Word segmentation can be performed based on rules such as regular expressions, or based on existing word segmentation tools;

[0071] Part-of-speech tagging: Automatically tag the word segmentation results with the corresponding part of speech for each word, such as noun, verb, adjective, etc.

[0072] Named entity recognition: Identify named entities in text and label them with corresponding entity types;

[0073] Syntactic analysis: Analyze the dependency relationship between words in text data and construct a syntactic dependency tree to facilitate further semantic analysis and processing.

[0074] In one embodiment, the feature extraction step in step 13 is specifically as follows:

[0075] Based on the text feature extraction model, sentiment analysis is performed on the text preprocessing data to obtain explicit sentiment text features, implicit sentiment text features and demand goals;

[0076] The acoustic features (spectral features, speech rate, intonation, pauses, volume, etc.) are extracted from the speech preprocessing data to obtain implicit emotional speech features.

[0077] Among them, the steps of building the text feature extraction model are:

[0078] Select an appropriate pre-trained model based on task requirements. In this embodiment, the BERT model is selected.

[0079] Adaptively adjust the BERT model based on the task objectives and build an initial text feature extraction model;

[0080] The initial text feature extraction model is iteratively trained using a training sample set constructed according to the task objective until the iteration stop condition is reached, thereby obtaining a text feature extraction model; wherein the training sample level includes an explicit emotion sample subset, an implicit emotion sample subset, and a demand target subset.

[0081] Among them, acoustic feature extraction is implemented based on the speech toolkit.

[0082] In one embodiment, the steps of acquiring the emotional state are:

[0083] The multi-head mutual attention mechanism is used to learn the interactive information between implicit emotional speech features and explicit emotional text features, and the interactive information between implicit emotional speech features and implicit emotional text features;

[0084] The cross-attention mechanism is used to learn the attention weights between the interactive information and the explicit sentiment text features to identify the sentiment state.

[0085] Furthermore, the data processing steps of the multi-head mutual attention mechanism are:

[0086] Obtain implicit emotional speech features, explicit emotional text features and implicit emotional text features;

[0087] Feature encoding of implicit emotional speech features, explicit emotional text features and implicit emotional text features;

[0088] The attention weights between different features are calculated based on the self-attention mechanism of each attention head in the multi-head mutual attention mechanism;

[0089] Then use the fusion strategy (weighted average, splicing, etc.) to fuse the learning results of each attention head to obtain the fusion features.

[0090] In this embodiment, the attention heads are split according to the needs. For example, in this embodiment, each head is focused on learning different types of emotional feature interactions.

[0091] Furthermore, the steps for obtaining the emotional state are:

[0092] The explicit sentiment text feature is used as the query (Query), and the fused features are the key (Key) and the value (Value). The dot product of the Query and each Key is calculated, and the softmax function is applied to obtain the attention weight. The obtained attention weight is used to perform weighted summation on the value (Value) in the fused feature to obtain the weighted fused feature. The weighted fused feature is input into the sentiment classifier to obtain the sentiment classification (positive, neutral, negative).

[0093] In one embodiment, the specific steps of determining the priority level are:

[0094] Step 21, quantify the demand target according to the preset emergency degree quantification table to obtain the demand quantification value;

[0095] Step 22: quantify the emotional state according to a preset emotional quantification table to obtain an emotional quantification value;

[0096] Step 23: Obtain a priority level using a priority rule based on the demand quantification value and the emotion quantification value.

[0097] In step 21, the urgency quantification table comprehensively considers the nature, scope of impact and urgency of solving different types of problems and divides them according to the types of problems. In this embodiment, they are divided into 5 categories, namely: common sense consultation (such as electricity usage common sense), business consultation (service channels, electricity usage policies, relevant regulations, etc.), electricity bill payment and bill objection, service interruption and restoration, and power grid security, with quantitative values ​​of 0.2, 0.4, 0.6, 0.8 and 1 respectively.

[0098] In step 22, the emotions are divided into five levels and assigned corresponding quantitative values. The more negative the emotion is, the larger the value is, as shown in Table 1.

[0099] Table 1

[0100]

[0101] In step 3, the priority rule is: determine whether the demand quantification value exceeds 0.8. If so, increase the priority to the highest priority. If not, calculate the demand quantification value and the emotion quantification value according to the weight value set based on historical data to determine their priority levels.

[0102] In one embodiment, the specific steps of demand forecasting are:

[0103] Step 31: Read the demand target, emotional state and historical interaction record of the voice data in order based on the priority level;

[0104] Step 32: Generate a demand forecast result using a demand forecast model according to the demand target, the emotional state and the historical interaction record.

[0105] In one embodiment, the steps for obtaining personalized reply content are:

[0106] Step 41: Call multiple pending response solutions according to the demand forecast result;

[0107] Step 42: calling a response template according to the emotional state;

[0108] Step 43: Use the demand target to select the to-be-responded plan, and generate a reply voice based on the to-be-responded template.

[0109] Furthermore, the response template can be regarded as an emotional constraint and a habit constraint. The emotional constraint includes tone, speaking speed, tone, etc., and the habit constraint includes hobbies, education level, age, etc.

[0110] In one embodiment, the above scheme is applied as follows: when a customer makes an electricity bill inquiry request, the voice command is first converted into text, and the query keywords and intents are accurately identified, such as "query this month's electricity bill", "historical payment records", etc. Based on the identified query keywords and intents, the customer's query needs are clarified, and the type and time range of electricity bill information to be queried are determined; at the same time, the customer's voice tone, speaking speed and other characteristics are analyzed to identify the customer's emotional state.

[0111] The customer's historical query records are retrieved from the database, including query time, query content, query results, etc. By analyzing the historical query records, the customer's electricity query behavior pattern is identified, such as query frequency, query time preference, query content preference, etc., and based on the historical query records, the customer's electricity usage habits can be inferred, such as whether they regularly query electricity bills, whether they pay attention to electricity usage in specific time periods, etc. According to the characteristics of the query content and historical query records, feature information is extracted from the current query content, historical query records and emotional state recognition results as input to the demand forecasting model to make real-time forecasts of the customer's query needs. For example, in this embodiment, the forecast results may include the type of electricity bill information that the customer may be concerned about, the time range, possible problems or electricity bill discounts, etc. Based on the forecast results, relevant electricity bill information or suggestions are presented to the customer. For example, if it is predicted that the customer needs to inquire about this month's electricity bill, the system will call up this month's electricity bill; if it is predicted that the customer may have questions about the electricity bill, the electricity bill explanation or complaint channel information will be called up; if there is a discount on the electricity bill, the relevant policy analysis will be called up. If the customer's next query information is about questions about the electricity bill, the electricity bill explanation content will be converted into the expression type that the customer likes, and played to the customer at a suitable speed and tone.

[0112] This embodiment also discloses an intelligent customer service question and answer optimization system based on sentiment analysis, such as Figure 2 As shown, it includes: a feature extraction module, which is used to read voice data and perform sentiment analysis and feature extraction on the voice data;

[0113] A priority level determination module, used to determine the priority level of the speech data according to the extracted features;

[0114] A demand prediction module, used to predict the demand for voice data in order based on priority;

[0115] The reply module is used to generate personalized reply content based on extracted features, predicted needs and historical interaction records.

[0116] In one embodiment, a system optimization module is also included, specifically: a service evaluation report is generated according to the collected feedback data and sentiment change trends, and the system is optimized based on the service evaluation report.

[0117] Furthermore, a service satisfaction evaluation report is prepared based on the customer's feedback data and the changing trends of the customer's emotional characteristics during the question-and-answer process, and the system is optimized and adjusted based on the service satisfaction evaluation report.

[0118] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0119] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent customer service question and answer optimization method based on sentiment analysis, characterized in that: The specific steps are: Reading voice data, and performing sentiment analysis and feature extraction on the voice data; Determining the priority level of the voice data based on the extracted features; Performing demand forecasting on the voice data in order based on the priority levels; Generate personalized reply content based on the extracted features, predicted needs and historical interaction records.

2. According to claim 1, the intelligent customer service question and answer optimization method based on sentiment analysis is characterized in that: The specific steps of sentiment analysis and feature extraction are: Recognize the read voice data to generate text data; Preprocessing the speech data and the text data to obtain speech preprocessing data and text preprocessing data; Extracting features from the speech preprocessing data and the text preprocessing data, wherein the extracted features include displayed emotion features, implicit emotion features, and demand targets; wherein the displayed emotion features are displayed emotion text features, and the implicit emotion features include implicit emotion speech features and implicit emotion text features; The displayed emotion feature and the implicit emotion feature are fused to obtain the emotion state.

3. According to claim 2, the intelligent customer service question and answer optimization method based on sentiment analysis is characterized in that: The preprocessing of the speech data includes speech denoising, silence detection, speech sentence segmentation and normalization.

4. According to claim 2, the intelligent customer service question and answer optimization method based on sentiment analysis is characterized in that: The preprocessing of the text data includes text denoising, text segmentation, part-of-speech tagging, named entity recognition and syntactic analysis.

5. According to claim 2, the intelligent customer service question and answer optimization method based on sentiment analysis is characterized in that: The steps of acquiring the emotional state are: The multi-head mutual attention mechanism is used to learn the interactive information between implicit emotional speech features and explicit emotional text features, and the interactive information between implicit emotional speech features and implicit emotional text features; The cross-attention mechanism is used to learn the attention weights between the interaction information and the explicit emotional text features to identify the emotional state.

6. According to claim 2, the intelligent customer service question and answer optimization method based on sentiment analysis is characterized in that: The specific steps of the priority level determination are: Quantify the demand target according to a preset urgency quantification table to obtain a demand quantification value; Quantifying the emotional state according to a preset emotional quantification table to obtain an emotional quantification value; The priority level is obtained by using a priority rule based on the demand quantification value and the emotion quantification value.

7. The intelligent customer service question and answer optimization method based on sentiment analysis according to claim 2 is characterized in that: The specific steps of demand forecasting are: Read the demand target, the emotional state, and the historical interaction record of the voice data in order based on the priority level; A demand prediction result is generated using a demand prediction model according to the demand target, the emotional state and the historical interaction record.

8. The intelligent customer service question and answer optimization method based on sentiment analysis according to claim 7 is characterized in that: The steps to obtain personalized reply content are as follows: Calling multiple pending response plans according to the demand forecast results; Invoking a response template according to the emotional state; The solution to be responded is selected using the demand target, and a reply voice is generated from the solution to be responded according to the template to be responded.

9. An intelligent customer service question and answer optimization system based on sentiment analysis, characterized in that: include: A feature extraction module is used to read voice data and perform sentiment analysis and feature extraction on the voice data; A priority level determination module, used for determining the priority level of the speech data according to the extracted features; A demand prediction module, used for performing demand prediction on the voice data in order based on the priority levels; The reply module is used to generate personalized reply content based on the extracted features, predicted needs and historical interaction records.

10. The intelligent customer service question and answer optimization system based on sentiment analysis according to claim 9, characterized in that: It also includes a system optimization module, specifically: generating a service evaluation report according to the collected feedback data and sentiment change trends, and optimizing the system based on the service evaluation report.

Images