User satisfaction analysis method and system based on artificial intelligence technology

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a user satisfaction analysis method based on an artificial intelligence technology, comprising the following steps: S1, loading data: loading the data including original outbound voice data and corresponding questionnaire data; s2, data preprocessing: converting the outbound voice data into a text, segmenting the converted text into a plurality of question and answer contexts based on the questionnaire data, performing error correction on the converted text based on the questionnaire data, and segmenting the outbound voice data into a plurality of corresponding voice segments; s3, performing satisfaction analysis: performing intelligent analysis on the segmented question and answer context and voice segments to obtain a satisfaction analysis result; and S4, outputting a satisfaction analysis result. The technical scheme of the invention has the advantages of no need of a large amount of manual analysis, high processing efficiency and low analysis cost, and is optimized based on the combination of speech recognition, natural language processing, large language model and multi-modal fusion model technologies and the combination of subdivided technical scenes and business characteristics.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a user satisfaction analysis method and system based on artificial intelligence technology. Background Art

[0002] In order to improve products and services and enhance user experience, companies need to continuously conduct user satisfaction research to understand user satisfaction with various indicators. A traditional user satisfaction research method is to ask users to rate their satisfaction with various services through manual outbound calls; then analyze the outbound voice data to obtain satisfaction research results. Traditional outbound voice data analysis is conducted manually, with manual listening and recording of various satisfaction scores. This method is inefficient, has huge manpower and training costs, and high quality inspection costs; and there are errors and deviations caused by personnel quality and subjectivity, resulting in unreliable analysis results. With the continuous advancement of AI technologies such as ASR (speech recognition), natural language processing, and large language models, the use of AI technology for automatic intelligent analysis has great social value.

[0003] In view of the shortcomings and deficiencies of traditional methods, the present invention proposes to automatically perform intelligent satisfaction analysis based on AI technology. The advantages of the present invention are that it can greatly reduce labor costs, improve efficiency, and ensure the stability of processing results. It can be applied to intelligent outbound calls and outbound call quality inspection scenarios in satisfaction research. Summary of the invention

[0004] The present invention provides a user satisfaction analysis method and system based on artificial intelligence technology to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, in a first aspect, the present invention provides the following technical solutions: A user satisfaction analysis method based on artificial intelligence technology includes the following steps: S1: Loading data: The loaded data includes the original outbound voice data and the corresponding questionnaire data; S2: Data preprocessing: converting the outbound voice data into text, dividing the converted text into multiple question-answer contexts based on the questionnaire data, correcting the converted text based on the questionnaire data, and dividing the outbound voice data into multiple corresponding voice segments; S3: Satisfaction analysis: intelligently analyze the segmented question and answer context and voice segments to obtain satisfaction analysis results; S4: Output satisfaction analysis results.

[0006] Preferably, step S1 specifically includes: S11, loading corresponding outbound call voice data based on the analyzed outbound call number; S12. Load corresponding questionnaire data based on the analyzed outbound call number.

[0007] Preferably, the questionnaire data includes a plurality of sub-items, each of which includes: a related question, a question type, and a value range of candidate answers to the question.

[0008] Preferably, step S2 specifically includes: S21, converting the outbound voice data into a text format; S22, segmenting the converted text according to each sub-item in the questionnaire data, each sub-item corresponding to a question and answer context; S23, removing noise and correcting errors from the converted text based on the questionnaire data; S24. Segment the outbound voice data according to the positions of the multiple question-answer contexts in the outbound voice data to obtain multiple voice segments corresponding to the multiple question-answer contexts.

[0009] Preferably, segmenting the converted text in step S22 includes: Based on the questions in the questionnaire data, the converted text is segmented through semantic matching combined with sequence matching; Among them, semantic matching includes performing semantic calculations on questions in the questionnaire data and the converted text data to identify the location of each question; Combined sequential matching includes correcting the matching errors according to the order of questions in the questionnaire data; The step S23 of removing noise and correcting errors from the converted text based on the questionnaire data includes: Use the questions in the questionnaire data to correct incorrectly converted questions, and use the value range of candidate answers in the questionnaire data to correct incorrectly converted answers.

[0010] Preferably, the satisfaction analysis in step S3 includes: Satisfaction analysis is performed on all sub-items, and the corresponding question and answer context and voice clips are analyzed to identify whether the user is satisfied with the sub-item. The satisfaction analysis results include "satisfied", "average", and "unsatisfied".

[0011] Preferably, step S3 specifically includes: S31. Use natural language processing technology to achieve satisfaction score recognition and obtain user rating data; S32. Construct a multimodal fusion model. The multimodal fusion model uses multiple encoders to obtain the representation information of the multimodal output respectively, and uses a neural network model to fuse them to obtain the output result, which is the satisfaction analysis result.

[0012] Preferably, the multiple encoders include a speech encoder, a text encoder, and a scoring encoder. The audio of the user's answer, the text description of the user's answer, and the user's scoring data in the annotation database are respectively input into the speech encoder, the text encoder, and the scoring encoder to obtain representation information of the multimodal output.

[0013] Preferably, the representation information output by multiple encoders is concatenated into a vector, which is converted into a vector of length 3 using a fully connected network, which is the satisfaction analysis result.

[0014] In a second aspect, the present invention provides a user satisfaction analysis system based on artificial intelligence technology, which is applicable to the user satisfaction analysis method based on artificial intelligence technology as described above, including: Data loading module: used to load the original outbound voice data and the corresponding questionnaire data; Data preprocessing module: used to convert outbound voice data into text, divide the converted text into multiple question-answer contexts based on questionnaire data, and divide the outbound voice data into corresponding multiple voice segments; Intelligent analysis module: including scoring recognition submodule and satisfaction recognition submodule. The scoring recognition submodule realizes satisfaction scoring recognition through natural language processing technology to obtain user rating data; the satisfaction recognition submodule uses a multimodal fusion model to use multiple encoders to obtain the representation information of multimodal outputs, and obtains the output result after fusion using a neural network model; The output module is used to output the satisfaction analysis results.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention automatically performs satisfaction analysis on the data of outbound questionnaire surveys based on artificial intelligence technology. The technical solution of the present invention has the advantages of not requiring a large amount of manual analysis, high processing efficiency and low analysis cost. It is based on the combination of speech recognition, natural language processing, large language model and multimodal fusion model technology, and optimizes the subdivided technical scenarios in combination with business characteristics, with a higher recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a user satisfaction analysis method based on artificial intelligence technology provided by an embodiment of the present invention; Figure 2 A schematic diagram of a data preprocessing process in a user satisfaction analysis method based on artificial intelligence technology provided by an embodiment of the present invention; Figure 3 A structural block diagram of a user satisfaction analysis system based on artificial intelligence technology provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0017] 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.

[0018] The embodiment of the present invention provides a method for analyzing user satisfaction based on artificial intelligence technology, comprising the following steps: S1: Loading data: The loaded data includes the original outbound voice data and the corresponding questionnaire data; S2: Data preprocessing: converting the outbound voice data into text, dividing the converted text into multiple question-answer contexts based on the questionnaire data, correcting the converted text based on the questionnaire data, and dividing the outbound voice data into multiple corresponding voice segments; S3: Satisfaction analysis: intelligently analyze the segmented question and answer context and voice segments to obtain satisfaction analysis results; S4: Output satisfaction analysis results.

[0019] Specifically, in one embodiment of the present invention, step S1 specifically includes: S11, loading corresponding outbound call voice data based on the analyzed outbound call number; S12. Load corresponding questionnaire data based on the analyzed outbound call number.

[0020] The questionnaire data includes multiple sub-items, each of which includes: related questions, question types, and the value range of candidate answers to the question.

[0021] For example, each sub-item is a certain attribute of a product or service, and the related question (the preset question that the user is required to answer) is to ask the user to rate the satisfaction with the attribute. For example, for the attribute "Internet access quality of broadband service", the related question is "Are you satisfied with the fixed broadband service of the telecommunications company? Please use an integer between one and ten to rate very satisfied (ten points), very dissatisfied (one point), how many points would you give?" The question type is a scoring question, and the range of candidate answers is one to ten.

[0022] Specifically, in one embodiment of the present invention, step S2 specifically includes: S21, converting the outbound voice data into a text format; S22, segmenting the converted text according to each sub-item in the questionnaire data, each sub-item corresponding to a question and answer context; S23, removing noise and correcting errors from the converted text based on the questionnaire data; S24. Segment the outbound voice data according to the positions of the multiple question-answer contexts in the outbound voice data to obtain multiple voice segments corresponding to the multiple question-answer contexts.

[0023] Specifically, the conversion of the outbound voice data into text format in step S21 is achieved by using ASR (speech recognition). The original outbound voice data is converted into text format, and the location information of each character is included, that is, the specific seconds that each character appears in the outbound voice data.

[0024] In one embodiment of the present invention, segmenting the converted text in step S22 includes: Based on the questions in the questionnaire data, the converted text is segmented by semantic matching combined with sequential matching. Segmenting the converted text questionnaire data into multiple question-answer contexts means that the interview content revolves around several sub-items, and each sub-item corresponds to a question-answer context in the text data.

[0025] Among them, semantic matching refers to performing semantic calculations on the questions in the questionnaire data and the converted text data to identify the location of each question. Combined with the sequential matching method, it refers to correcting the matching errors according to the order of the questions in the questionnaire data.

[0026] For example, the implementation method of semantic matching is: first, use the text vectorization model to convert all text fragments into corresponding semantic vectors; then, use the text vectorization model to convert all questions in the questionnaire data into corresponding semantic vectors; finally, calculate the similarity between the semantic vector of each question and the semantic vector of the text fragment, and the one with the highest similarity is the matching result; the result of semantic matching is the context fragment corresponding to each question; The implementation method of sequential matching is that the questions in the questionnaire data have a sequence, and the matching results must also conform to this sequence. If the matching results are inconsistent with the sequence of the questions in the questionnaire data, the results are corrected according to the calculation method of the edit distance. Among them, correcting the results according to the calculation method of the edit distance means that among multiple correction schemes, the method that requires the least operation and the least number of corrections is used as the final result. Specifically, when correcting a certain matching item, if the matching result with the highest similarity is inconsistent with the sequential matching, first obtain the result that is consistent with the sequential matching result from other matching results as a candidate item; then select the matching item with the highest similarity among the candidate items as the final correction result.

[0027] Specifically, removing noise and correcting errors from the converted text in step S21 refers to correcting erroneous data based on the real questions and the value range of candidate answers in the questionnaire data. The speech recognition process easily recognizes the text as an erroneous homophone. If there is a difference between the question and the question in the questionnaire data, the question in the questionnaire data is directly used for correction.

[0028] In one embodiment of the present invention, the segmentation of the outbound voice data in step S24 includes: The outbound voice data is segmented based on the text segmentation results. In the voice recognition results, each character corresponds to the position information in the voice file. According to the text segmentation results, the position of the beginning and end of each segment in the outbound voice data can be obtained. The outbound voice data is segmented according to the beginning and end positions to obtain voice segments. In one embodiment of the present invention, the satisfaction analysis in step S3 includes: Satisfaction analysis is performed on all sub-items, and the corresponding question and answer context and voice clips are analyzed to identify whether the user is satisfied with the sub-item. The satisfaction analysis results include "satisfied", "average", and "unsatisfied".

[0029] Specifically, in one embodiment of the present invention, step S3 specifically includes: S31. Use natural language processing technology to achieve satisfaction score recognition and obtain user rating data; S32. Construct a multimodal fusion model. The multimodal fusion model uses multiple encoders to obtain the representation information of the multimodal output respectively, and uses a neural network model to fuse them to obtain the output result, which is the satisfaction analysis result.

[0030] Specifically, the satisfaction score recognition by natural language processing technology in step S31 specifically includes: Implementation method 1: Input the question-answering context and analysis target into the large language model through prompt words. The implementation of a prompt word is "The text enclosed in three backticks is an interview dialogue, and the text enclosed in two backticks is the question of concern. Now you need to perform the following analysis based on the content of the interview dialogue and the question of concern: Please help determine the score of the interviewee for the question of concern? If no actual user score number is found, answer unknown"; Due to the limitations of the large language model, its output has a certain degree of uncertainty. After the large language model is recognized, the result needs to be verified. If it is not a legal scoring result, try to correct the recognition result. If it is still not a legal result after correction, it is considered that there is no valid score in the context and the scoring result is "unknown". Regular expressions are used to verify the recognition results of the large language model, and the large language model is used to correct the irregular scoring results. Correcting irregular scoring results using a large language model can be achieved by setting prompt words. For example, a prompt word can be implemented as "The following may be some irregular text. Please convert it into an Arabic numeral and return it. If it cannot be converted into a number, return unknown."

[0031] Implementation method 2: Treat the scoring recognition task as a supervised text classification task and implement it by training a text classification model. First, build a labeled database, and then use the labeled database to train the classification model. The output category of the trained classification model is the scoring result.

[0032] In one embodiment of the present invention, the multiple encoders include a speech encoder, a text encoder, and a scoring encoder. The audio of the user's answer, the text description of the user's answer, and the user's scoring data in the annotation database are respectively input into the speech encoder, the text encoder, and the scoring encoder to obtain representation information of the multimodal output.

[0033] Among them, the speech encoder and text encoder are implemented using existing pre-trained models. The rating encoder is implemented by encoding the ratings, converting the input user rating data into a vector model. The training process of the multimodal fusion model is carried out through multiple stages. In the first stage, the language model and audio model are trained separately through pre-training. The training steps in this stage are optional steps. If an existing encoder is used, this stage can be skipped. In the second stage, the fusion model structure is trained. During training, the encoder model weights are frozen, and only the fusion model structure is trained. One embodiment of the scoring encoder is to use a positional encoding method of a transformer model, wherein the input is a number between 1 and 10, and the output is a vector of fixed dimension; Another embodiment of the scoring encoder is to use a one-hot encoding method, where the input is a number between 1 and 10 and the output is a vector of fixed dimension. In one embodiment of the present invention, the representation information output by multiple encoders is concatenated into a vector, which is converted into a vector of length 3 using a fully connected network, which is the satisfaction analysis result.

[0034] The embodiment of the present invention further provides a user satisfaction analysis system based on artificial intelligence technology, which is applicable to the user satisfaction analysis method based on artificial intelligence technology as described above, including: Data loading module: used to load the original outbound voice data and the corresponding questionnaire data. Supported loading sources include local files, databases, and cloud storage methods; Data preprocessing module: used to convert outbound voice data into text, divide the converted text into multiple question-answer contexts based on questionnaire data, and divide the outbound voice data into corresponding multiple voice segments; The outbound voice data to text conversion is realized through the speech recognition model, which converts the original outbound voice data into text format, including the location information of each word, that is, the specific seconds each word appears in the outbound voice data; The converted text questionnaire data is divided into multiple question-answer contexts, which means that the interview content is carried out around several sub-items, and each sub-item corresponds to a question-answer context in the text data; The outbound voice data is divided into a plurality of corresponding voice segments, which means that the outbound voice data is divided based on the text segmentation result; in the voice recognition result, each character corresponds to the position information in the voice file, and according to the text segmentation result, the position of the beginning and the end of each segment in the outbound voice data can be obtained, and the outbound voice data is divided according to the position of the beginning and the end to obtain the voice segments; Intelligent analysis module: including scoring recognition submodule and satisfaction recognition submodule. The scoring recognition submodule realizes satisfaction scoring recognition through natural language processing technology to obtain user rating data; the satisfaction recognition submodule uses a multimodal fusion model to use multiple encoders to obtain the representation information of multimodal outputs, and obtains the output result after fusion using a neural network model; The output module is used to output the satisfaction analysis results. The output module supports returning the analysis results directly through the API, writing them to local disk files, and storing them in the database.

[0035] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A user satisfaction analysis method based on artificial intelligence technology, characterized in that: The following steps are involved: S1: Loading data: The loaded data includes the original outbound voice data and the corresponding questionnaire data; S2: Data preprocessing: converting the outbound voice data into text, dividing the converted text into multiple question-answer contexts based on the questionnaire data, correcting the converted text based on the questionnaire data, and dividing the outbound voice data into multiple corresponding voice segments; S3: Satisfaction analysis: intelligently analyze the segmented question and answer context and voice segments to obtain satisfaction analysis results; S4: Output satisfaction analysis results.

2. The user satisfaction analysis method based on artificial intelligence technology according to claim 1 is characterized in that: Step S1 specifically includes: S11, loading corresponding outbound call voice data based on the analyzed outbound call number; S12. Load corresponding questionnaire data based on the analyzed outbound call number.

3. The user satisfaction analysis method based on artificial intelligence technology according to claim 2 is characterized in that: The questionnaire data includes multiple sub-items, each of which includes: related questions, question types, and the value range of candidate answers to the question.

4. The user satisfaction analysis method based on artificial intelligence technology according to claim 1 is characterized in that: Step S2 specifically includes: S21, converting the outbound voice data into a text format; S22, segmenting the converted text according to each sub-item in the questionnaire data, each sub-item corresponding to a question and answer context; S23, removing noise and correcting errors from the converted text based on the questionnaire data; S24. Segment the outbound voice data according to the positions of the multiple question-answer contexts in the outbound voice data to obtain multiple voice segments corresponding to the multiple question-answer contexts.

5. The user satisfaction analysis method based on artificial intelligence technology according to claim 4 is characterized in that: The segmentation of the converted text in step S22 includes: Based on the questions in the questionnaire data, the converted text is segmented through semantic matching combined with sequential matching; Among them, semantic matching includes performing semantic calculations on questions in the questionnaire data and the converted text data to identify the location of each question; Combined sequential matching includes correcting the matching errors according to the order of questions in the questionnaire data; The step S23 of removing noise and correcting errors from the converted text based on the questionnaire data includes: Use the questions in the questionnaire data to correct incorrectly converted questions, and use the value range of candidate answers in the questionnaire data to correct incorrectly converted answers.

6. The user satisfaction analysis method based on artificial intelligence technology according to claim 3 is characterized in that: The satisfaction analysis described in step S3 includes: Satisfaction analysis is performed on all sub-items, and the corresponding question and answer context and voice clips are analyzed to identify whether the user is satisfied with the sub-item. The satisfaction analysis results include "satisfied", "average", and "unsatisfied".

7. The user satisfaction analysis method based on artificial intelligence technology according to claim 6 is characterized in that: Step S3 specifically includes: S31. Use natural language processing technology to achieve satisfaction score recognition and obtain user rating data; S32. Construct a multimodal fusion model. The multimodal fusion model uses multiple encoders to obtain the representation information of the multimodal output respectively, and uses a neural network model to fuse them to obtain the output result, which is the satisfaction analysis result.

8. The user satisfaction analysis method based on artificial intelligence technology according to claim 7 is characterized in that: The multiple encoders include a speech encoder, a text encoder, and a scoring encoder. The audio of the user's answer, the text description of the user's answer, and the user's scoring data in the annotation database are respectively input into the speech encoder, the text encoder, and the scoring encoder to obtain representation information of the multimodal output.

9. The method for analyzing user satisfaction based on artificial intelligence technology according to claim 7, characterized in that: The representation information output by multiple encoders is concatenated into a vector and converted into a vector of length 3 using a fully connected network, which is the satisfaction analysis result.

10. A user satisfaction analysis system based on artificial intelligence technology, characterized in that: The method for analyzing user satisfaction based on artificial intelligence technology according to any one of claims 1 to 9 comprises: Data loading module: used to load the original outbound voice data and the corresponding questionnaire data; Data preprocessing module: used to convert outbound voice data into text, divide the converted text into multiple question-answer contexts based on questionnaire data, and divide the outbound voice data into corresponding multiple voice segments; Intelligent analysis module: including scoring recognition submodule and satisfaction recognition submodule. The scoring recognition submodule realizes satisfaction scoring recognition through natural language processing technology to obtain user rating data; the satisfaction recognition submodule uses a multimodal fusion model to use multiple encoders to obtain the representation information of multimodal outputs, and obtains the output result after fusion using a neural network model; The output module is used to output the satisfaction analysis results.

Images