A Business and Government Affairs Dialogue Management Method and System Based on Sentiment Analysis

Abstract

The present application discloses a business and government affairs dialogue management method and system based on sentiment analysis. The method includes: when a user enters the home page of the business and government affairs online dialogue system, selects a consultation question category and conducts an online consultation, collecting the dialogue text data and interaction behavior data of the user in the online consultation dialogue; using a dialogue sentiment state recognition model to identify the sentiment state category in the dialogue text data; determining the dialogue interaction intention of the user based on the sentiment state and interaction behavior data; according to the dialogue interaction intention, if it is determined that the reply content of the user needs to be adjusted, adjusting the content reply strategy for the user in the current online consultation dialogue; if it is determined that the reply content of the user does not need to be adjusted, continuing to have a dialogue with the user according to the established content in the current online consultation dialogue. The present application can provide corresponding reply content recommendations according to the user's sentiment state, improve the user interaction efficiency, and enhance the user's experience and satisfaction with the business dialogue system.

Description

Technical Field

[0001] This application relates to the technical field of intelligent services, and particularly to a business and government affairs dialogue management method and system. Background Art

[0002] Technologies such as artificial intelligence and big data have developed rapidly. The combination of virtual economy and real economy driven by the Internet and big data has brought revolutionary changes to people's work and life styles. Among them, the human-computer dialogue interaction in different task scenarios has broken the traditional interaction mode, and the data-driven interaction algorithm has introduced personalized factors such as interaction content, emotional changes, and user behavior data into the human-computer dialogue interaction. By analyzing the user's interaction intention through artificial intelligence algorithms, corresponding content recommendations in the interaction dialogue are given, thereby improving the user's interaction efficiency.

[0003] The human-computer dialogue process generally refers to the process of a user communicating and interacting with government, merchants, or other organizational personnel in a specific scenario. This dialogue process takes place in a specific context and involves the interaction between the user and government, merchants, or organizational personnel. Its purposes may include information query, problem solving, service acquisition, etc. In the scenario of business and government affairs, government organizational personnel can usually provide corresponding interaction feedback by observing the user's dialogue tone and facial emotion changes. However, in the human-computer interaction process of business and government affairs dialogue, the interaction system often fails to capture the user's facial emotion changes for detailed analysis. The user's emotional fluctuations during the interaction may have a certain impact on the fluency and efficiency of the interaction. Unfortunately, these online interaction scenarios often ignore the impact of the user's emotional changes in a specific environment and fail to provide corresponding content recommendations according to the user's emotions. Summary of the Invention

[0004] This application provides a business and government affairs dialogue management method and system based on emotion analysis, aiming to solve the technical problem that corresponding content recommendations are not provided according to the user's emotions in the existing online dialogue interaction of business and government affairs.

[0005] In a first aspect, a business and government affairs dialogue management method based on emotion analysis includes:

[0006] S1, when the user enters the homepage of the business and government affairs online dialogue system, selects the consultation question category and conducts online consultation, collect the user's dialogue text data and interaction behavior data in the online consultation dialogue through the user online interaction interface;

[0007] S2, use the dialogue emotion state recognition model to identify the emotion state category in the dialogue text data;

[0008] S3, based on the emotion state and interaction behavior data, determine the user's dialogue interaction intention;

[0009] S4. Based on the dialogue interaction intention, determine whether it is necessary to adjust the content of the user's reply in the current online consultation dialogue;

[0010] S5. If it is determined that it is necessary to adjust the content of the user's reply, adjust the content reply strategy for the user in the current online consultation dialogue; if it is determined that it is not necessary to adjust the content of the user's reply, continue to have a dialogue with the user according to the established content in the current online consultation dialogue.

[0011] In the above solution, optionally, step S2 includes:

[0012] Use a dialogue semantic feature extraction model to extract the dialogue semantic features in the dialogue text data;

[0013] Use a dialogue emotion feature extraction model to extract the dialogue emotion features in the dialogue text data;

[0014] Fuse the dialogue semantic features and dialogue emotion features in a dynamic weighting manner to obtain fused features;

[0015] Identify the emotion state category according to the fused features.

[0016] In the above solution, further optionally, the dialogue semantic feature extraction model adopts a large language pre-training and fine-tuning model, and the training process of the dialogue semantic feature extraction model includes:

[0017] Collect a publicly available dialogue interaction data set;

[0018] Use the publicly available dialogue interaction data set to train a large language model;

[0019] Collect dialogue data in the industrial and commercial administrative service scenarios, and preprocess the dialogue data;

[0020] Use the preprocessed dialogue data to fine-tune the large language model to obtain a dialogue semantic feature extraction model;

[0021] Evaluate the performance of the dialogue semantic feature extraction model.

[0022] In the above solution, further optionally, the dialogue emotion feature extraction model adopts a bidirectional autoencoder model, and the training process of the dialogue emotion feature extraction model includes:

[0023] Collect a publicly available emotion dialogue data set;

[0024] Preprocess the emotion dialogue data set;

[0025] Construct the structure of the bidirectional autoencoder model;

[0026] Fine-tune the bidirectional autoencoder model using the preprocessed sentiment dialogue dataset to obtain a dialogue sentiment feature extraction model;

[0027] Evaluate the performance of the dialogue sentiment feature extraction model.

[0028] In the above solution, further optionally, the fusion of the dialogue semantic features and dialogue sentiment features by dynamic weighting to obtain fusion features includes:

[0029] Concatenate the dialogue semantic features and dialogue sentiment features to obtain a first sub-fusion feature;

[0030] Perform dynamic multiplication of the dialogue semantic features and the first sub-fusion feature to obtain a second sub-fusion feature;

[0031] Perform dynamic multiplication of the dialogue sentiment features and the first sub-fusion feature to obtain a third sub-fusion feature;

[0032] Concatenate the second sub-fusion feature and the third sub-fusion feature to obtain a fusion feature.

[0033] In the above solution, optionally, the sentiment state category is a negative emotion or a positive emotion, the interaction behavior data includes the duration of browsing and consulting the dialogue and the operation frequency of the interaction device, and the dialogue interaction intention is to be interested or not interested in the current reply content.

[0034] In the above solution, further optionally, step S3 includes:

[0035] When the sentiment state category is a negative emotion, and the duration of browsing and consulting the dialogue is lower than the preset duration threshold and the operation frequency of the interaction device is higher than the preset frequency threshold, it is determined that the user's dialogue interaction intention is not interested in the current reply content;

[0036] When the sentiment state category is a positive emotion, and the duration of browsing and consulting the dialogue is higher than the preset duration threshold and the operation frequency of the interaction device is lower than the preset frequency threshold, it is determined that the user's dialogue interaction intention is interested in the current reply content.

[0037] In the above solution, further optionally, step S4 includes:

[0038] If the user's dialogue interaction intention is not interested in the current reply content, it is determined that the user's reply content needs to be adjusted;

[0039] If the user's dialogue interaction intention is interested in the current reply content, it is determined that the user's reply content does not need to be adjusted.

[0040] Optionally, in the above solution, after step S5, the method further includes:

[0041] S6. Saving the entire online consultation dialogue and the changing trend of the user's emotional state category in the entire online consultation dialogue into the database.

[0042] In a second aspect, a business and government affairs dialogue management system based on sentiment analysis includes:

[0043] A dialogue interaction data collection module, configured to collect the user's dialogue text data and interaction behavior data in the online consultation dialogue through the user online interaction interface when the user enters the homepage of the business and government affairs online dialogue system, selects the consultation question category, and conducts an online consultation;

[0044] An emotional state recognition module, configured to recognize the emotional state category in the dialogue text data by using a dialogue emotional state recognition model;

[0045] A dialogue interaction intention understanding module, configured to determine the user's dialogue interaction intention according to the emotional state and interaction behavior data;

[0046] A judgment module, configured to judge whether it is necessary to adjust the reply content to the user in the current online consultation dialogue according to the dialogue interaction intention;

[0047] A dialogue content adjustment module, configured to, if it is determined that it is necessary to adjust the reply content to the user, adjust the content reply strategy to the user in the current online consultation dialogue; if it is determined that it is not necessary to adjust the reply content to the user, continue to have a dialogue with the user according to the established content in the current online consultation dialogue.

[0048] Compared with the prior art, the present application has at least the following beneficial effects:

[0049] In the business and government affairs dialogue management method based on sentiment analysis provided in the embodiment of the present application, by using the dialogue interaction data to analyze the change of the user's emotional state in the dialogue environment, and then analyzing the user's interaction intention according to different emotional states and the content of human-computer interaction, and then giving corresponding answer recommendations according to the interaction intention; in the business and government affairs online dialogue interaction, the emotional change of the user can be automatically captured, the intention of the user can be accurately understood, so as to provide more targeted replies and services, and provide reply content more in line with the user's needs; realize providing corresponding reply content recommendations according to the user's emotional state in the business and government affairs online dialogue interaction, improve the user interaction efficiency, and at the same time enable better satisfaction of the user's emotional needs, and enhance the user's experience and satisfaction with the business and government affairs dialogue system. Description of the Drawings

[0050] Figure 1Schematic flowchart of a business and government affairs dialogue management method based on sentiment analysis provided in an embodiment of the present application;

[0051] Figure 2 Schematic flowchart in a sentiment recognition model for feature fusion based on a dynamic vision self-attention model in an embodiment of the present application;

[0052] Figure 3 Another schematic flowchart of a business and government affairs dialogue management method based on sentiment analysis provided in an embodiment of the present application;

[0053] Figure 4 Block diagram of the module architecture of a business and government affairs dialogue management system based on sentiment analysis provided in an embodiment of the present application;

[0054] Figure 5 Schematic diagram of the system framework of a business and government affairs dialogue management system based on sentiment analysis provided in an embodiment of the present application;

[0055] Figure 6 Schematic flowchart of the processing in a dialogue intention understanding model provided in an embodiment of the present application;

[0056] Figure 7 Schematic flowchart of a dialogue question and answer adjustment process based on the emotional state provided in an embodiment of the present application. Detailed implementation manners

[0057] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0058] In the description of the present application: Unless otherwise specified, "a plurality of" means two or more. Terms such as "first", "second", "third", etc. in the present application are intended to distinguish the objects being referred to, and do not have special significance in terms of technical connotations (for example, it should not be understood as emphasizing the importance level or order, etc.). Expressions such as "including", "comprising", "having", etc. also mean "not limited to" (certain units, components, materials, steps, etc.).

[0059] In one embodiment, as Figure 1 shown, a business and government affairs dialogue management method based on sentiment analysis is provided, and the method includes the following steps:

[0060] S1, when the user enters the home page of the business and government affairs online dialogue system, selects the category of the consultation question and conducts an online consultation, collect the dialogue text data and interaction behavior data of the user in the online consultation dialogue through the user online interaction interface.

[0061] Among them, the online dialogue system presents the user's online dialogue web page, including display interfaces such as user login, the system homepage showing dialogue scenario categories, and the dialogue web page displaying dialogue content. The interaction behavior data includes the duration of browsing and consulting dialogues and the operation frequency of the interaction device.

[0062] In step S1, dialogue interaction data is collected. The interaction information between the user and the system is collected through the user's online interaction interface, mainly for obtaining the consultation text data of the dialogue interaction in the user's industrial and commercial administrative environment.

[0063] The interaction text data and interaction behavior data in the user's online dialogue interaction process are collected through the online dialogue system, which is used to analyze the emotional state changes and interaction intention understanding in the user's interaction process.

[0064] S2. Using the dialogue emotional state recognition model, identify the emotional state categories in the dialogue text data.

[0065] Among them, the emotional state categories are negative emotions or positive emotions.

[0066] In the online dialogue emotion recognition model, the emotional changes contained in the dialogue statements in the user's online dialogue interaction process are a feedback on the degree of interest in the current consultation content, which has a certain universality.

[0067] Furthermore, step S2 includes:

[0068] S21. Using the dialogue semantic feature extraction model, extract the dialogue semantic features in the dialogue text data;

[0069] S22. Using the dialogue emotion feature extraction model, extract the dialogue emotion features in the dialogue text data;

[0070] S23. Fuse the dialogue semantic features and dialogue emotion features in a dynamically weighted manner to obtain fused features;

[0071] S24. According to the fused features, identify the emotional state categories.

[0072] Specifically, in step S21, the dialogue semantic feature extraction model adopts a large language pre-training and fine-tuning model. The training process of the dialogue semantic feature extraction model includes:

[0073] Collect a publicly available dialogue interaction data set;

[0074] Use the publicly available dialogue interaction data set to train a large language model;

[0075] Collect dialogue data in the industrial and commercial administrative scenario, and preprocess the dialogue data;

[0076] Fine-tune the large language model using the preprocessed dialogue data to obtain a dialogue semantic feature extraction model;

[0077] Evaluate the performance of the dialogue semantic feature extraction model.

[0078] That is to say, when training the dialogue semantic feature extraction model, it is necessary to collect dialogue interaction data in the industrial and commercial administrative field, including users' questions and system answers, which will be used as the training data for the dialogue semantic feature extraction model. During the data collection process, it is necessary to ensure the quality and diversity of the data to improve the generalization ability of the model.

[0079] For the dialogue data in the industrial and commercial administrative field, it is necessary to select a suitable large language model (such as BERT, GPT, etc.). These pre-trained models are usually pre-trained on large-scale text data and have good semantic understanding and generation capabilities.

[0080] Load the selected language model into the training environment and fine-tune it using the dialogue data in the industrial and commercial administrative field. The purpose of fine-tuning is to enable the language model to better understand and encode the dialogue semantics in the industrial and commercial administrative field to meet the extraction requirements of the dialogue semantic features in the industrial and commercial administrative field. The fine-tuning process usually includes setting hyperparameters, selecting a suitable optimizer, defining a loss function, etc. This step is to enable the language model to better understand and encode the dialogue semantics in the industrial and commercial administrative field.

[0081] After the pre-training fine-tuning is completed, use the test set to evaluate the model to evaluate its performance on the dialogue data in the industrial and commercial administrative field. Metrics such as accuracy and F1 value can be used to evaluate the quality of the model.

[0082] Generally speaking, for dialogue semantic feature extraction, a large language model is pre-trained and fine-tuned on the data collected in the industrial and commercial administrative field to adapt to the high-level semantic encoding in downstream tasks. The dialogue text data in the user's online interaction dialogue is encoded and input into the fine-tuned large language model to extract the high-dimensional semantic features of the image.

[0083] Specifically, in step S22, the dialogue emotion feature extraction model uses a bidirectional autoencoder model. The training process of the dialogue emotion feature extraction model includes:

[0084] Collect publicly available emotion dialogue datasets;

[0085] Preprocess the emotion dialogue datasets;

[0086] Construct the structure of the bidirectional autoencoder model;

[0087] Fine-tune the bidirectional autoencoder model using the preprocessed emotion dialogue datasets to obtain a dialogue emotion feature extraction model;

[0088] Perform performance evaluation on the dialogue emotion feature extraction model.

[0089] That is to say, when training the dialogue emotion feature extraction model, it is first necessary to collect a publicly available emotion dialogue dataset, and these data will be used as the training data for the BERT model. During the data collection process, it is necessary to ensure the quality and diversity of the data to improve the generalization ability of the model. Preprocess the collected raw data, including word segmentation, stop word removal, emotion type annotation, etc.

[0090] Use a pre-trained BERT model (such as BERT-base, BERT-large, etc.) as the base model to construct an emotion feature extraction model. Open-source toolkits such as the Transformers library of Hugging Face or the Keras API of TensorFlow can be used to construct the BERT model.

[0091] Based on the BERT model, use the publicly available emotion dialogue dataset for fine-tuning to adapt to the dialogue context in the industrial and commercial administrative fields. The fine-tuning process usually includes selecting an appropriate optimizer, setting hyperparameters, training the model, etc. After the model fine-tuning is completed, it is necessary to evaluate the model using the test set. Metrics such as accuracy, precision, recall, and F1 value are usually used to evaluate the performance of the model.

[0092] Generally speaking, regarding the impact of the emotional state change of the user's dialogue process in the industrial and commercial administrative scenario on the service, train a bidirectional encoder representation model based on self-attention on the publicly available emotion dialogue dataset, and then use the trained model to extract the emotional features in the dialogue.

[0093] Specifically, in step S23, the dialogue semantic features and dialogue emotion features are fused by a dynamic weighting method to obtain fused features, including:

[0094] Concatenate the dialogue semantic features and dialogue emotion features to obtain the first sub-fused feature;

[0095] Dynamically multiply the dialogue semantic features with the first sub-fused feature to obtain the second sub-fused feature;

[0096] Dynamically multiply the dialogue emotion features with the first sub-fused feature to obtain the third sub-fused feature;

[0097] Concatenate the second sub-fused feature and the third sub-fused feature to obtain the fused feature.

[0098] Among them, splicing features refers to combining two features together according to certain rules to form a new feature vector, which jointly participates in dialogue classification or other tasks in the model. Specifically, the two features are concatenated element by element in a certain way or weighted and summed according to a certain ratio to obtain a new feature vector.

[0099] The first sub-fusion feature is a new feature obtained by splicing the dialogue semantic feature and the dialogue emotion feature. This feature vector contains the semantic information and emotion information of the dialogue. The second sub-fusion feature is a new feature obtained by dynamically multiplying the dialogue semantic feature and the first sub-fusion feature; this operation can be used to strengthen or weaken the weight of the dialogue semantic feature in the overall feature, so as to better express the relationship between semantics and emotion. Similarly, the third sub-fusion feature is a new feature obtained by dynamically multiplying the dialogue emotion feature and the first sub-fusion feature; this operation can be used to strengthen or weaken the weight of the dialogue emotion feature in the overall feature and further adjust the relationship between semantics and emotion. The second sub-fusion feature and the third sub-fusion feature are spliced to obtain the final fusion feature. This fusion feature contains the semantics, emotion of the dialogue and the relationship between them, and can be used for subsequent classification, analysis or decision-making tasks.

[0100] By splicing and dynamically multiplying different features, semantic and emotion information can be fully fused, improving the model's understanding and expression ability of dialogue content, so as to better meet user needs.

[0101] Figure 2 It is the processing flow in the emotion recognition model based on the dynamic vision self-attention model in the embodiments of the present application. In other words, aiming at the influence of the user's emotional changes on the understanding of dialogue intentions during the dialogue interaction process, the present application proposes a dynamic vision self-attention model to dynamically fuse the high-dimensional semantic features extracted by the large language pre-training fine-tuning model and the emotion change features extracted by the bidirectional auto-encoding model of the self-attention mechanism.

[0102] In dialogue emotion recognition, the dynamic vision self-attention model proposed in the embodiments of the present application can encode and add the high-dimensional semantic features of the dialogue to the bidirectional self-attention encoding network to enhance the discriminability of the emotion change features.

[0103] In the dynamic attention fusion model, first, the high-dimensional semantic encoding features are copied to the same dimension as the emotion change features, and then multiplied by the output of the multi-head self-attention mechanism model. The dynamic attention fusion model incorporates the high-dimensional semantic features into the feature extraction process of the dialogue emotion change by the method of dynamic multiplication to increase the discriminability of the emotion features.

[0104] The output of the multi-head self-attention mechanism model is a set of high-dimensional semantic encoding features, which are formed by concatenating the outputs of multiple attention heads. In the multi-head self-attention mechanism, for each attention head, the model learns a different set of attention weights to calculate the importance of each position in the input sequence for other positions, and then extracts the key information in the sequence. In this way, the addition of multiple attention heads can make the model have stronger expressive power and generalization ability.

[0105] Dynamically multiplying two features means that in the feature fusion process, the two features are multiplied element-wise to obtain a new feature vector. The "dynamic" here means that during the multiplication process, the model will strengthen or suppress different parts according to the specific situation, so as to improve the discrimination of the model for emotional change features.

[0106] During the dynamic multiplication process, the dimensionality sizes of the two features need to be the same. Specifically, one of the feature vectors can be copied multiple times to make its dimensionality size the same as that of the other feature vector, and then element-wise multiplication is performed. In this way, the similar parts in the two features will be strengthened, while the irrelevant parts will be suppressed, thereby increasing the discrimination of the model for emotional features.

[0107] In the dynamic attention fusion model, dynamically multiplying the high-dimensional semantic encoding features and the emotional change features can enable the model to better combine semantic features and emotional features, thereby improving the accuracy of dialogue emotion recognition.

[0108] Generally speaking, aiming at the influence of the emotional state change on the interaction process in the industrial and commercial government affairs dialogue interaction process, through the dynamic weighted method based on the dynamic visual self-attention model, the features of different dimensions are hierarchically fused to classify the current user's emotional state.

[0109] In other words, the online dialogue emotion recognition model in step S2 includes three modules: dialogue semantic feature extraction, dialogue emotion feature extraction, dialogue feature fusion and emotion state recognition:

[0110] (1) The dialogue emotion feature extraction model uses a large number of publicly available text emotion datasets to train the corresponding text autoencoder model to extract the emotion features of the dialogue interaction image data, and inputs the preprocessed data such as encoding, word segmentation, and normalization of the dialogue interaction data collected under the online dialogue platform into the dialogue interaction encoding model to extract the user's emotion features.

[0111] (2) Semantic feature extraction during the dialogue interaction process. For the specific tasks of users in different interaction scenarios, a publicly available dialogue interaction data is also used to train a feature extraction model. Different from the emotional features, semantic features are more universal and require a large-scale professional scenario dataset for training. Therefore, a bidirectional auto-encoding network model is used to extract semantic features.

[0112] (3) Dialogue feature fusion. Due to the particularity of dialogue statements for different scenario tasks, various fusion strategies at the feature layer and model layer need to be considered to eliminate the interference of different task scenarios. Therefore, online dialogue feature fusion considers the historical dialogue information, and at the same time fuses the semantic features and emotional features of the historical information through dynamic weighting, and comprehensively analyzes the effectiveness of different hierarchical features for emotion recognition.

[0113] S3. Determine the user's dialogue interaction intention based on the emotional state and interaction behavior data.

[0114] Among them, the dialogue interaction intention is to be interested or not interested in the current reply content.

[0115] Dialogue interaction intention understanding. Judge the user's interaction intention based on the recognized dialogue emotional state and interaction features.

[0116] Further, step S3 includes:

[0117] When the emotional state category is a negative emotion, and the browsing duration of the consultation dialogue is lower than the preset duration threshold, and the operation frequency of the interaction device is higher than the preset frequency threshold, it is determined that the user's dialogue interaction intention is not interested in the current reply content;

[0118] When the emotional state category is a positive emotion, and the browsing duration of the consultation dialogue is higher than the preset duration threshold, and the operation frequency of the interaction device is lower than the preset frequency threshold, it is determined that the user's dialogue interaction intention is interested in the current reply content.

[0119] In other words, step S3 conducts dialogue interaction intention understanding. When the user's dialogue emotion is a negative emotion category, and at the same time the user's interaction dialogue statements may be shorter and the interaction frequency is higher, at this time the user is not interested in the dialogue content. And when the user is in a positive emotional state, the user's interaction process statements may be longer, the interaction frequency is lower, and the browsing time of the interaction content is longer. At this time, the user is interested in the interaction content.

[0120] S4. Judge whether it is necessary to adjust the user's reply content in the current online consultation dialogue according to the dialogue interaction intention.

[0121] Further, step S4 includes:

[0122] If the user's dialogue interaction intention is not interested in the current reply content, it is determined that the user's reply content needs to be adjusted;

[0123] If the user's dialogue interaction intention is interested in the current reply content, it is determined that the user's reply content does not need to be adjusted.

[0124] S5. If it is determined that the user's reply content needs to be adjusted, adjust the content reply strategy for the user in the current online consultation dialogue; if it is determined that the user's reply content does not need to be adjusted, continue to have a dialogue with the user according to the established content in the current online consultation dialogue.

[0125] The dialogue content adjustment dynamically adjusts the dialogue consultation recommendation strategy for the user according to the system's judgment of the user's interaction intention, improving the user interaction efficiency.

[0126] In other words, in step S5, the dialogue content is adjusted, and corresponding interaction feedback is carried out according to the emotional characteristics of the shopping user and their interaction frequency. When the user's emotion is in a positive state during the interaction, and at the same time the interaction frequency with the dialogue system is low and the time for browsing the dialogue content is long, introduce similar topics to the user. When the user shows a negative emotional state, with a high interaction frequency and a short time for browsing the dialogue content, adjust the dialogue content strategy for the user.

[0127] Further, after step S5, it further includes:

[0128] S6. Save the entire online consultation dialogue and the change trend of the user's emotional state category in the entire online consultation dialogue to the database.

[0129] For the working process schematic diagram of the industrial and commercial administrative dialogue management method provided by the embodiments of the present application, reference can also be made to Figure 3 。

[0130] First, the industrial and commercial administrative dialogue interaction system collects the user's dialogue data during the consultation process and the user's behavior data during the interaction process. Then the user enters the home page of the system to browse and select the consultation question category of interest, and then selects the most interesting consultation question to enter the consultation page within a specific time.

[0131] During this process, the system conducts real-time analysis on the collected dialogue data through the dialogue emotional state recognition model, continuously monitoring the user's consultation process.

[0132] When the system detects that the user's emotional state continuously changes to a negative emotional category, it adjusts the conversation content to match the user's consultation content to satisfy the user. When the user's emotional state is positive, the system predicts the user's consultation intention based on the user's interaction data and continuously recommends similar conversation content; when the user is satisfied with the recommended consultation answer, the current round of consultation stage ends. Otherwise, the conversation content is adjusted until the consultation answer satisfies the user, and the system operation ends.

[0133] In the embodiments of the present application, in view of the lack of dialogue emotion analysis in current industrial and commercial administration affairs, a dialogue emotion recognition method based on the combination of multi-dimensional emotion features is proposed. By analyzing the dialogue interaction data, the change of the user's emotional state in the dialogue environment is analyzed, and the user's interaction intention is analyzed according to different emotional states and the content of human-computer interaction, so as to improve the user's interaction experience.

[0134] In view of the lack of analysis of the influence of emotional factors on interaction intention in current online dialogue interaction in industrial and commercial administration affairs, the embodiments of the present application propose a dialogue management method for industrial and commercial administration affairs based on emotional semantic analysis.

[0135] In view of the above problems, based on the dialogue interaction data in the industrial and commercial administration dialogue management system, the embodiments of the present application deeply explore the processing method of interaction data with the characteristics of emotional feature changes in the dialogue process for the text data characteristics in the user's dialogue interaction perception and cognition, so as to achieve robust emotion recognition, accurate intention understanding and human-computer dialogue interaction, and finally realize the integration of the interaction feedback process, providing new means and new ways to break through the development bottleneck of the artificial intelligence and human-computer interaction industries.

[0136] In one embodiment, as Figure 4 shown, a dialogue management system for industrial and commercial administration affairs based on emotion analysis is provided, including the following program modules:

[0137] A dialogue interaction data acquisition module 401, configured to collect the user's dialogue text data and interaction behavior data in the online consultation dialogue through the user online interaction interface when the user enters the home page of the industrial and commercial administration dialogue system, selects a consultation question category, and conducts an online consultation;

[0138] An emotional state recognition module 402, configured to recognize the emotional state category in the dialogue text data by using a dialogue emotional state recognition model;

[0139] A dialogue interaction intention understanding module 403, configured to determine the user's dialogue interaction intention according to the emotional state and interaction behavior data;

[0140] A judgment module 404, configured to judge whether it is necessary to adjust the reply content to the user in the current online consultation dialogue according to the dialogue interaction intention;

[0141] The dialogue content adjustment module 405 is used to adjust the content reply strategy to the user in the current online consultation dialogue if it is determined that the user's reply content needs to be adjusted; if it is determined that the user's reply content does not need to be adjusted, continue to have a dialogue with the user according to the established content in the current online consultation dialogue.

[0142] For the specific implementation content of each module, reference can be made to the definition of a business and government affairs dialogue management method based on sentiment analysis in the above text, which will not be elaborated here.

[0143] The system provided by the embodiments of the present application includes two modules: semantic sentiment analysis and dialogue intention understanding. The semantic sentiment analysis module includes: dialogue data collection, emotional semantic feature analysis, and emotional state recognition; the dialogue intention understanding module includes: dialogue system interface, influence analysis of emotional semantics on dialogue intention, and dialogue intention understanding.

[0144] The system analyzes the user's emotional state by collecting the dialogue data of the user during the interaction process in the business and government affairs environment, predicts the dialogue intention according to the user's emotional state, and gives corresponding answer recommendations, so as to improve the user's interaction efficiency. The method of the present invention can comprehensively verify the influence on the dialogue intention based on recognition in the business and government affairs environment and evaluate the dialogue understanding ability of the system.

[0145] It can also be said that the business and government affairs dialogue management system includes the following modules:

[0146] (1) Dialogue system login module.

[0147] This module is used by end users, and the interaction behavior data and dialogue data collection during the user's online business and government affairs dialogue interaction occur in this module.

[0148] The user registers and logs in to the system by entering the username and password. The user logs in to the account and fills in the basic information and inputs it into the user database for storage. After successful login, enter the personal space, and the user's basic personal information can be seen.

[0149] (2) Dialogue interaction system module.

[0150] In the online dialogue interaction system, several consultation field categories will be displayed on the home page. The user selects the field category they want to consult, enters the detailed introduction after selecting the field, and then conducts a consultation dialogue under the corresponding field category questions. The real-time dialogue data of the user is collected during the dialogue process.

[0151] (3) Dialogue intention understanding model.

[0152] During the conversation, the user's conversation semantics, emotional state, browsing time, and interaction frequency are a kind of feedback on the degree of interest in the current conversation content, which has a certain universality.

[0153] When the user is not interested in the conversation content during the interaction process, the user's emotion is usually negative, and at the same time, the user's browsing time is short and the interaction frequency is high. When the user is interested in the conversation commodity, the user's emotional state is usually more positive. At this time, the user's browsing time is long and the interaction frequency is low, and more attention is paid to the conversation content itself of the interaction.

[0154] (4) Interactive conversation content adjustment module.

[0155] Make corresponding interactive feedback according to the user's emotional characteristics and their interaction frequency.

[0156] When the user's emotion is in a positive state during the conversation, and at the same time the interaction frequency with the conversation system is low and the browsing time is long, the system determines that the user is satisfied with the reply to the conversation content of the interaction, and the system continues the conversation according to the established content.

[0157] When the user shows a negative facial emotional state, with a high interaction frequency and a short browsing time, the system determines that the user is not satisfied with the reply to the conversation content of the interaction, and the system adjusts the conversation content.

[0158] This module dynamically adjusts the conversation content, so that the user's consultation content experience always stays in the conversation part that the user is interested in, reduces the user's conversation process, completes the user's consultation process as soon as possible, reduces time waste, and realizes the intelligent interaction between online users and the industrial and commercial administrative dialogue management system.

[0159] On the other hand, for the system framework schematic diagram of this industrial and commercial administrative dialogue management system, reference can also be made to Figure 5 , this dialogue management system includes a system layer, a data layer, a feature layer, an emotion layer, an intention layer, and an interaction layer:

[0160] (1) The system layer is the industrial and commercial administrative interaction system. The user registers and logs in to the system by entering the user name and password. The user logs in to the account and fills in the basic information and inputs it into the user database for storage. After successful login, the user enters the personal space and can see the user's basic personal information.

[0161] (2) The data layer is for the system to collect the conversation text data and interaction behavior data during the user's online interaction process.

[0162] The dialogue interaction data collection module is mainly used to collect the semantic and emotional features of the text during users' online consultations and judge the emotional state of the current user. The input device's online interaction behavior data collection uses an online consultation platform to collect behavior data such as the duration of users' browsing, consulting, and answering, and the operation frequency of the interaction device, which is mainly used to analyze the degree of interest of users in the system dialogue during the online consultation process.

[0163] (3) The feature layer includes text semantic feature extraction, emotional feature extraction, and interaction data feature extraction.

[0164] Among them, for the semantic feature extraction of dialogue text data, a large language model is fine-tuned and trained in the industrial and commercial administrative service scenarios to extract the high-level semantic features of the dialogue text; for the emotional change state of users during the consultation process, a bidirectional encoder representation model based on self-attention is trained on the publicly available emotional dialogue dataset, and then the trained model is used to extract the emotional features in the dialogue; for the feature extraction of interaction data, statistical features such as the median, mean, minimum, maximum, range, standard deviation, and variance of the interaction data are extracted.

[0165] (4) The emotion layer extracts the high-level semantic features in the dialogue through the fine-tuning of the large language model and extracts the emotional change features in the dialogue through the self-attention bidirectional encoder model.

[0166] Regarding the impact of the emotional state change on the interaction process during the industrial and commercial administrative service dialogue interaction, different dimensions of features are hierarchically fused in a dynamically weighted manner through a dynamic vision self-attention model to classify the current user's emotional state.

[0167] (5) The intention layer discriminates the degree of interest in the current consultation dialogue content through the user's dialogue emotional state, browsing time, and interaction frequency during the shopping process.

[0168] When the user's emotion is in the negative emotion category, and at the same time the user's browsing system answer time is short and the interaction frequency is high, the user is not interested in the reply content of the consultation at this time. When the user is in a positive emotional state, the user's browsing system answer time is long and the interaction frequency is low, and the user is interested in the reply content of the consultation at this time.

[0169] (6) The interaction layer is a dialogue consultation answer adjustment module, which judges whether the system changes the consultation answer content according to the evaluation result of the online dialogue consultation system on the degree of user interest.

[0170] The dialogue consultation recommendation adjustment module mainly relies on the system's recognition of dialogue emotion and understanding of dialogue intention to make a judgment. When the result of dialogue emotion recognition is a negative emotional state, the system determines that the user's degree of interest is in a declining state at this time, and then adjusts the consultation answer to the user to improve the user's consultation efficiency.

[0171] In one embodiment, as Figure 6 shown, a schematic diagram of the processing flow of a dialogue intention understanding model is provided.

[0172] When the user registers, a questionnaire is used to survey the user's past consulting tendencies in the industrial and commercial fields. For example, whether the user is more inclined to company registration consulting or share transfer policy consulting, and dialogue tendency characteristics are formed according to the questionnaire answers filled in by the user.

[0173] After the user enters the system homepage and selects a consulting category, the dialogue interaction data of the user is collected. Each piece of dialogue data is input into a large language preprocessing fine-tuning model and a self-attention bidirectional autoencoder model pre-trained on a dialogue dataset and an emotion dataset to obtain the high-dimensional semantic features and emotion change features of the dialogue statements. Through a dynamic self-attention fusion model, the high-dimensional semantic features are integrated into the extraction process of the emotion change features in a dynamically weighted form to enhance the discriminability of the emotion features.

[0174] At the same time, feature processing is performed on the interaction time and interaction frequency of the interaction behavior data collected by the system, and statistical features such as the median, mean, minimum value, maximum value, range, standard deviation, and variance of the interaction data are extracted as interaction features.

[0175] Based on the user's shopping tendency characteristics, emotion change characteristics, high-dimensional semantic characteristics, and interaction behavior characteristics, weighted fusion is performed to judge the user's dialogue intention.

[0176] When the user's expression is negative, the user is not interested in the interaction content. At this time, the user's browsing time of the system answer is short and the interaction frequency is high.

[0177] When the user's expression is usually more positive, the user is interested in the interaction content. At this time, the user's browsing time of the system answer is long and the interaction frequency is low.

[0178] In one embodiment, as Figure 7 shown, a schematic diagram of the dialogue question and answer adjustment process based on the emotional state is provided.

[0179] The industrial and commercial government affairs online dialogue system needs to complete the function of providing answers to the consulting content to the user and timely analyzing and storing the selection of the consulting question category, so that the user can conduct efficient question consulting. When the user enters the dialogue system homepage and selects a consulting question category, the system starts to collect the dialogue interaction data of the user by analyzing the current emotional state of the user.

[0180] When the user shows a negative state, the system believes that the answer given to the user is not satisfactory, and at this time the system will change the answer to the question raised by the user; when the user shows a positive emotional state, the system believes that the answer to the user's question is the best answer, and continues the dialogue Q&A process, while saving the user's current emotional content. When the user is satisfied with the answer to the question raised, but the conversation has not ended yet, the continuous conversation process continues until the conversation is completed.

[0181] After the user finishes learning, the system will automatically save the emotional change trend during the entire online conversation process, which is convenient for the system to query the conversation status of the user at any time period, understand the consultation content that the user is interested in, so as to recommend better dialogue Q&A for the user subsequently.

[0182] In one embodiment, a computer-readable storage medium is further provided, on which a computer program is stored, covering all or part of the processes in the above-mentioned embodiment methods.

[0183] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

Claims

1. A method for managing business and government affairs dialogue based on sentiment analysis, characterized in that: include: S1, when a user enters the homepage of the online dialogue system for industrial and commercial government affairs, selects a consultation question category and conducts online consultation, the user's dialogue text data and interaction behavior data in the online consultation dialogue are collected through the user's online interaction interface; S2, using a dialogue emotion state recognition model to identify the emotion state category in the dialogue text data; S3, determining the user's dialogue interaction intention based on the emotional state and interaction behavior data; S4, judging whether it is necessary to adjust the user's reply content in the current online consultation dialogue according to the dialogue interaction intention; S5, if it is determined that the user's reply content needs to be adjusted, adjusting the content reply strategy for the user in the current online consultation dialogue; If it is determined that there is no need to adjust the user's reply content, continue to communicate with the user according to the established content in the current online consultation dialogue; Step S2 includes: Extracting the conversation semantic features from the conversation text data using a conversation semantic feature extraction model; Extracting the conversation emotion features from the conversation text data using a conversation emotion feature extraction model; The conversation semantic features and the conversation emotional features are fused by means of dynamic weighting to obtain fused features; identifying the emotional state category according to the fused features; The conversation semantic features and conversation emotional features are fused in a dynamic weighted manner to obtain fused features, including: The conversation semantic feature and the conversation emotional feature are concatenated to obtain a first sub-fusion feature; Dynamically multiplying the conversation semantic feature with the first sub-fusion feature to obtain a second sub-fusion feature; Dynamically multiplying the conversation emotion feature with the first sub-fusion feature to obtain a third sub-fusion feature; The second sub-fusion feature and the third sub-fusion feature are concatenated to obtain a fusion feature.

2. The method for managing business and government affairs dialogue based on sentiment analysis according to claim 1 is characterized in that: The dialogue semantic feature extraction model adopts a large language pre-training fine-tuning model, and the training process of the dialogue semantic feature extraction model includes: Collecting public conversational interaction datasets; Use public conversational interaction datasets to train large language models; Collecting conversation data in business and government affairs scenarios and preprocessing the conversation data; Using the preprocessed conversation data to fine-tune the large language model to obtain a conversation semantic feature extraction model; The performance of the conversation semantic feature extraction model is evaluated.

3. The method for managing business and government affairs dialogue based on sentiment analysis according to claim 1 is characterized in that: The dialogue emotion feature extraction model adopts a bidirectional autoencoder model, and the training process of the dialogue emotion feature extraction model includes: Collect public emotional dialogue datasets; Preprocessing the emotional dialogue dataset; Construct the structure of a bidirectional autoencoder model; Using the preprocessed emotional dialogue dataset to fine-tune the bidirectional autoencoder model to obtain a dialogue emotional feature extraction model; The performance of the conversation emotion feature extraction model is evaluated.

4. The method for managing business and government affairs dialogue based on sentiment analysis according to claim 1 is characterized in that: The emotional state category is negative emotion or positive emotion, the interactive behavior data includes the duration of browsing consultation dialogue and the operation frequency of interactive devices, and the dialogue interaction intention is to be interested in the current reply content or not interested in the current reply content.

5. The method for managing business and government affairs dialogue based on sentiment analysis according to claim 4 is characterized in that: Step S3 includes: When the emotional state category is negative emotion, and the duration of the browsing consultation dialogue is less than the preset duration threshold, and the operation frequency of the interactive device is higher than the preset frequency threshold, it is determined that the user's dialogue interaction intention is not interested in the current reply content; When the emotional state category is positive emotion, and the duration of the browsing consultation dialogue is higher than the preset duration threshold, and the operation frequency of the interactive device is lower than the preset frequency threshold, it is determined that the user's dialogue interaction intention is to be interested in the current reply content.

6. The method for managing business and government affairs dialogue based on sentiment analysis according to claim 4 is characterized in that: Step S4 includes: If the user's conversational interaction intention is that he is not interested in the current reply content, it is determined that the user's reply content needs to be adjusted; If the user's dialogue interaction intention is to be interested in the current reply content, it is determined that there is no need to adjust the user's reply content.

7. The method for managing business and government affairs dialogue based on sentiment analysis according to claim 1 is characterized in that: After step S5, the method further comprises: S6, saving the entire online consultation dialogue and the changing trend of the user's emotional state category in the entire online consultation dialogue into a database.

8. A business and government dialogue management system based on sentiment analysis, characterized in that: include: The dialogue interaction data collection module is used to collect the dialogue text data and interaction behavior data of the user in the online consultation dialogue through the user online interaction interface when the user enters the homepage of the industrial and commercial government affairs online dialogue system, selects the consultation question category and conducts online consultation; An emotional state recognition module, used to recognize the emotional state category in the conversation text data using a conversation emotional state recognition model; A dialogue interaction intention understanding module, used to determine the user's dialogue interaction intention based on the emotional state and interaction behavior data; A judgment module, used to judge whether it is necessary to adjust the user's reply content in the current online consultation dialogue according to the dialogue interaction intention; A dialogue content adjustment module is used to adjust the content response strategy for the user in the current online consultation dialogue if it is determined that the user's response content needs to be adjusted; If it is determined that there is no need to adjust the user's reply content, continue to communicate with the user according to the established content in the current online consultation dialogue; The emotional state recognition module includes: Extracting the conversation semantic features from the conversation text data using a conversation semantic feature extraction model; Extracting the conversation emotion features from the conversation text data using a conversation emotion feature extraction model; The conversation semantic features and the conversation emotional features are fused by means of dynamic weighting to obtain fused features; identifying the emotional state category according to the fused features; The conversation semantic features and conversation emotional features are fused in a dynamic weighted manner to obtain fused features, including: The conversation semantic feature and the conversation emotional feature are concatenated to obtain a first sub-fusion feature; Dynamically multiplying the conversation semantic feature with the first sub-fusion feature to obtain a second sub-fusion feature; Dynamically multiplying the conversation emotion feature with the first sub-fusion feature to obtain a third sub-fusion feature; The second sub-fusion feature and the third sub-fusion feature are concatenated to obtain a fusion feature.

Images