An artificial intelligence-based interaction system and method

Abstract

The present invention provides an artificial intelligence-based interaction system and method, including six modules: user intention recognition, historical data storage, feature extraction, intelligent decision-making, content generation, and feedback optimization. The system analyzes user input, extracts features, makes intelligent decisions to generate interaction content, and optimizes performance based on user feedback. The present invention can achieve efficient and accurate user intention recognition and response generation, improving the accuracy of interaction and user satisfaction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. More specifically, the present invention relates to an interaction system and method based on artificial intelligence. Background Art

[0002] In the field of artificial intelligence, with the development of technology, human-computer interaction has become more and more frequent and important. Most of the existing interaction systems rely on preset rules or simple pattern matching techniques, and these methods are often unable to cope when dealing with complex and variable user intentions. In addition, when processing natural language input, these systems often cannot accurately understand the true needs of users, resulting in low interaction efficiency. In the decision-making process, traditional algorithms may not be able to make full use of context information and historical interaction data, thus affecting the accuracy and personalization of decisions. In terms of content generation, existing systems often lack flexibility and creativity and are difficult to generate interaction content that meets user expectations. Finally, most systems lack an effective feedback mechanism and cannot optimize and learn themselves according to user feedback to improve the interaction quality.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the accuracy of user intention recognition is insufficient, feature extraction and utilization are not sufficient, the intelligence level of the decision-making algorithm is limited, the personalization and creativity of content generation are insufficient, and the system feedback optimization mechanism is missing. Summary of the Invention

[0004] The present invention provides an interaction system and method based on artificial intelligence.

[0005] In the first aspect of the present invention, an interaction system based on artificial intelligence is provided, including:

[0006] A user intention recognition module: for parsing the user input content and recognizing the user intention.

[0007] A historical data storage module: storing the historical data of past human-computer interactions to provide a data basis for subsequent analysis.

[0008] A feature extraction module: extracting key features from the user input and related data for algorithm analysis.

[0009] An intelligent decision-making module: adopting an improved decision tree algorithm to make interaction decisions based on the extracted features.

[0010] A content generation module: generating corresponding interaction content according to the decision result of the intelligent decision-making module.

[0011] A feedback optimization module: collecting user feedback on the interaction content and optimizing the system performance.

[0012] Furthermore, the user intention recognition module includes the following steps:

[0013] Step 1: Segment the text input by the user to obtain a word sequence , where is the number of words, represents the word sequence and is the

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021] Furthermore, the historical data storage module stores the records of each interaction

[0019]

[0020] where represents the number of interactions, is the user input for the th interaction, is the user intention recognized for the th interaction, is the output of the system for the th interaction.

[0021] Furthermore, the feature extraction module includes the following steps:

[0022] Step 1: Extract the word frequency feature from the user input text The word frequency calculation formula is

[0023]

[0024] Among them, represents the word frequency in the user input text and is the number of times the word appears in the text . is the total number of times all words appear in the text .

[0025] Step 2: Combine the data in the historical data storage module and calculate the similarity feature between the user input and the historical input . The similarity calculation formula is

[0026]

[0027] Among them, represents the similarity between the user input and the historical input , and are respectively the word frequencies of the word in the user input and the historical input .

[0028] Furthermore, the improved decision tree algorithm of the intelligent decision-making module includes the following steps:

[0029] Step 1: Initialize the decision tree and use all features as candidate splitting features.

[0030] Step 2: For each candidate splitting feature, select the optimal splitting point according to the information gain ratio. The information gain ratio calculation formula is

[0031]

[0032] Among them, is the information gain ratio of the feature to the data set ,

[0033]

[0034] is the information gain

[0035]

[0036] is the information entropy of the data set , is the number of categories in the data set, is the data subset that belongs to the th category in the data set. is the number of samples in the dataset , is the dataset in which the feature takes the value of the sample subset

[0037]

[0038] is the eigenvalue of the feature , is the feature of all possible values

[0039] Step 3: Split the dataset according to the optimal splitting point to generate child nodes

[0040] Step 4: Recursively repeat Step 2 and Step 3 for each child node until the stopping condition is met. The stopping condition includes that the node data belongs to the same category or reaches the maximum depth

[0041] Step 5: Based on the generated decision tree, make a decision on the extracted features to obtain the interactive decision result

[0042] Furthermore, the content generation module includes the following steps

[0043] Step 1: Determine the type of interactive content to be generated according to the decision result of the intelligent decision module. The type of interactive content includes text and image

[0044] Step 2: If it is of text type, select a suitable template from the predefined text template library , combine the key information in the user input , and generate specific text content through the text filling algorithm . The text filling formula is , where represents the text filling function, which is used to fill the key information into the template to generate specific text content

[0045] Furthermore, the feedback optimization module includes the following steps

[0046] Step 1: Collect the feedback information of the user on the interactive content , and the feedback information includes the user satisfaction score . The user satisfaction score ranges from , and the text feedback content

[0047] ​​Step 2: According to the feedback information , if the user satisfaction score is lower than the set threshold, and the set threshold is , adjust the relevant decision tree model parameters, and the adjustment formula is

[0048]

[0049] wherein, is the adjusted parameter, is the parameter before adjustment, is the learning rate, is the gradient of the loss function calculated based on the feedback information with respect to the parameter .

[0050] Furthermore, during the training process of the intent classification neural network model, the cross-entropy loss function is adopted, and the loss function formula is

[0051]

[0052] wherein, is the cross-entropy loss value, is the number of training samples, is the total number of intent categories, is the sample belonging to the true label of the intent category , if the sample belongs to the intent category then is , otherwise it is , is the probability that the model predicts that the sample belongs to the intent category , is the word vector sequence corresponding to the sample .

[0053] Furthermore, during the decision tree generation process, a pre-pruning strategy is adopted. When the information gain ratio of the node is less than the set threshold , stop splitting the node.

[0054] In the second aspect of the present invention, an artificial intelligence-based interaction method is provided, including:

[0055] Parse the user input content and identify the user intent;

[0056] Extract key features from the user input and relevant data;

[0057] An improved decision tree algorithm is adopted to make interactive decisions based on the extracted features;

[0058] According to the decision result of the intelligent decision-making module, corresponding interactive content is generated;

[0059] Collect user feedback on the interactive content to optimize the system performance.

[0060] According to the above embodiments of the present invention, it has at least the following beneficial effects: The artificial intelligence interaction system of the present invention can improve the accuracy and efficiency of human-computer interaction by integrating a user intention recognition module, a historical data storage module, a feature extraction module, an intelligent decision-making module, a content generation module, and a feedback optimization module. The system uses a pre-trained word vector model and an intention classification neural network model to deeply analyze user input, accurately identify user intentions, and make intelligent decisions through an improved decision tree algorithm to generate interactive content that meets user needs. At the same time, the system can collect user feedback and dynamically optimize the decision tree model parameters to further improve the system performance.

[0061] In addition, the feature extraction module of the system can extract key features from user input and calculate the input similarity in combination with historical interaction data to provide rich context information for intelligent decision-making. The content generation module can select a suitable template from a predefined template library according to the decision result and generate specific text content in combination with the key information in user input to enhance the personalization and relevance of the interactive content. The feedback optimization mechanism of the system can adjust the model parameters according to the user satisfaction score and text feedback content to achieve self-learning and continuous optimization of the system, thereby improving user satisfaction and interaction experience. Brief Description of the Drawings

[0062] By referring to the accompanying drawings and reading the following detailed description, the above and other purposes, features, and advantages of the exemplary embodiments of the present invention will become easily understandable. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, where:

[0063] Figure 1 It is a schematic structural diagram of an artificial intelligence-based interaction system provided by an embodiment of the present invention;

[0064] Figure 2 It is a schematic flow diagram of an artificial intelligence-based interaction method provided by an embodiment of the present invention;

[0065] Figure 3 It schematically shows a schematic structural diagram of an electronic device according to an embodiment of the present invention. Detailed Description of the Embodiments

[0066] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and do not limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to be able to fully convey the scope of the present invention to those skilled in the art.

[0067] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, a device, a device, a method, or a computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: completely hardware, completely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0068] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0069] The following reference Figure 1 , Figure 1 is a schematic structural diagram of an artificial intelligence-based interaction system provided for an embodiment of the present invention. As Figure 1 shown, an artificial intelligence-based interaction system 100 includes:

[0070] A user intention recognition module 101: used to parse the user input content and recognize the user intention.

[0071] A historical data storage module 102: stores the historical data of past human-computer interactions and provides a data basis for subsequent analysis.

[0072] A feature extraction module 103: extracts key features from the user input and related data for algorithm analysis.

[0073] An intelligent decision-making module 104: adopts an improved decision tree algorithm to make interaction decisions based on the extracted features.

[0074] A content generation module 105: generates corresponding interaction content according to the decision result of the intelligent decision-making module.

[0075] A feedback optimization module 106: collects user feedback on the interaction content and optimizes the system performance.

[0076] It should be noted that the system includes a user intention recognition module, which is responsible for parsing the user input content and recognizing the user intention. The user intention recognition module is one of the core components of the system, and it determines the operation or information that the user wants to execute by analyzing the user input text. In this process, the user intention refers to the specific needs or goals expressed by the user through the input.

[0077] Specifically, the user intention recognition module includes performing word segmentation on the text input by the user to obtain a word sequence, where the word sequence is a set composed of a series of words, and each word represents a word or phrase in the text. Word segmentation is the process of splitting continuous text into discrete lexical units. Then, using a pre-trained word vector model, each word is mapped to a word vector. The word vector is the process of converting words in the text into numerical vectors, which enables the computer to process and analyze text data.

[0078] Preferably, when the user intention recognition module inputs the word vector sequence into the intention classification neural network model, it determines the user intention by calculating the probabilities belonging to each intention category. This probability calculation formula takes into account the scores of each intention category output by the neural network model and finally selects the intention category with the highest probability as the recognition result.

[0079] Furthermore, in this process, different neural network architectures and optimization algorithms can be adopted to improve the recognition accuracy. In addition, the system can also adapt to specific application scenarios by adjusting the parameters of the word vector model or selecting different pre-trained models.

[0080] In some embodiments, the user intention recognition module includes the following steps:

[0081] Step 1: The text input by the user is subjected to word segmentation to obtain a word sequence , where is the number of words, represents the word sequence and is the

[0082] Step 2: Using a pre-trained word vector model, each word in the word sequence is mapped to a word vector to obtain a word vector sequence .

[0083] Step 3: Input the word vector sequence into the intention classification neural network model, and calculate the probabilities belonging to each intention category

[0084]

[0085] through the formula , where represents the probability of belonging to the intention category given the word vector sequence , is the score corresponding to the intention category output by the intention classification neural network model The score, is the total number of intent categories, and the intent category with the highest final probability is the recognized user intent.

[0086] It should be noted that this module includes three steps: First, perform word segmentation on the text input by the user. Second, use a pre-trained word vector model to map each word to a word vector. Finally, input the word vector sequence into the intent classification neural network model to calculate the probabilities belonging to each intent category. Here, word segmentation refers to the process of splitting continuous natural language text into independent lexical units, and the pre-trained word vector model refers to a model pre-trained using a large amount of text data that can convert words into numerical vectors of a fixed dimension.

[0087] Specifically, in the word segmentation step, a word segmentation algorithm in a natural language processing library, such as jieba segmentation (for Chinese) or NLTK segmentation (for English), can be used to split the text input by the user into a word sequence. The pre-trained word vector model can be Word2Vec, GloVe, or BERT, etc. These models can convert each word into a vector in a high-dimensional space, retaining the semantic relationships between words. In the intent classification neural network model, structures such as a multi-layer perceptron (MLP) or a convolutional neural network (CNN) can be used. The input is the word vector sequence, and the output is the probabilities belonging to each intent category.

[0088] Preferably, for word segmentation, different word segmentation tools and parameter settings can be selected according to the characteristics of different languages and domains. For example, for Chinese text, the word segmentation algorithm can be set to identify the most appropriate lexical boundaries. For the pre-trained word vector model, a pre-trained model most relevant to the application domain can be selected, or the model can be fine-tuned as needed to better adapt to a specific dataset.

[0089] Furthermore, during the training process of the intent classification neural network model, cross-validation and hyperparameter optimization techniques can be used to determine the best network structure and parameters to improve the generalization ability of the model. In addition, ensemble learning methods, such as random forest or gradient boosting machine, can also be considered to further improve the accuracy of intent recognition.

[0090] In some embodiments, the historical data storage module stores the records of each interaction

[0091]

[0092] where, represents the number of interactions, is the user input for the th interaction, is the user intent recognized for the th interaction, is the output of the th interaction system.

[0093] It should be noted that this module is responsible for storing the records of each human-computer interaction, including user input, the recognized user intention, and the system output. Here, the interaction record refers to all the data generated during the communication between the system and the user, and these data are used to analyze user behavior and optimize system performance. The historical data storage module is an important part of the system, which saves all past interaction data and provides a basis for subsequent data analysis and feature extraction.

[0094] Specifically, the historical data storage module can be implemented using a database system, where the record of each interaction consists of a triple (user input, user intention, system output). User input refers to the information provided by the user during the interaction with the system, user intention is the intention of the user recognized by the system, and system output is the response generated by the system according to the user intention. The database can be a relational database such as MySQL or a non-relational database such as MongoDB, and the specific choice depends on the data structure and query requirements. In the database, a unique identifier can be set for each interaction record for easy retrieval and management.

[0095] Preferably, the historical data storage module can be further refined to include data encryption and backup mechanisms to ensure data security and reliability. For data encryption, algorithms such as AES or RSA can be used to protect the stored data from unauthorized access. The backup mechanism can regularly back up the database to prevent data loss or damage.

[0096] Furthermore, the function of automatic data cleaning and archiving can also be implemented to optimize the use of storage space. When storing data, data compression technology can also be considered to reduce the occupancy of storage space while ensuring fast data reading. For the selection of the database, in addition to traditional relational and non-relational databases, time series databases or graph databases can also be considered, especially when the interaction data has time series characteristics or complex association relationships.

[0097] In some embodiments, the feature extraction module includes the following steps:

[0098] Step 1: Extract the word frequency feature from the user input text The word frequency calculation formula is

[0099]

[0100] where represents the word in the user input text The word frequency, is the word The number of occurrences in the text is the total number of occurrences of all words in the text is the text in all words.

[0101] Step 2: Combine the data in the historical data storage module to calculate the user input and the historical input similarity feature, and the similarity calculation formula is

[0102]

[0103] where represents the similarity between the user input and the historical input , and are respectively the word frequencies of the word in the user input and the historical input .

[0104] It should be noted that this module involves extracting word frequency features from the user input text and calculating the similarity features between the user input and the historical input in combination with the data in the historical data storage module. Here, the word frequency feature refers to the number of occurrences of each word in the text, and the similarity feature is a measure of the similarity in vocabulary usage between the new user input and the historical input. These features are crucial for understanding the user's input content and identifying their intentions.

[0105] Specifically, the feature extraction module first extracts the word frequency feature from the user input text. The word frequency refers to the number of occurrences of a specific word in the text, which can be calculated by counting the occurrence frequency of each word. The calculation formula is: word frequency (word, text) = the number of occurrences of the word in the text / the total number of occurrences of all words in the text. Then, in combination with the data in the historical data storage module, the similarity feature between the user input and the historical input is calculated, which can be achieved by comparing the word frequencies in the new user input with the corresponding word frequencies in the historical data. The similarity calculation formula involves comparing the word frequencies of each word to calculate the similarity between the two.

[0106] Preferably, the operation steps of the feature extraction module can be further refined. For example, when extracting the word frequency feature, a threshold can be set to ignore words with extremely low occurrence frequencies to reduce the influence of noise. When calculating the similarity feature, more complex similarity calculation methods such as cosine similarity can be used, which can better capture the similarity between word frequency vectors.

[0107] Furthermore, machine learning methods such as random forest or support vector machine can be introduced to learn from these features and predict user intent. For similarity calculation, a machine learning model can be considered to identify and weight the importance of different words, so as to more accurately reflect the similarity between the user input and the historical input.

[0108] In some embodiments, the improved decision tree algorithm of the intelligent decision module includes the following steps:

[0109] Step 1: Initialize the decision tree and use all features as candidate splitting features.

[0110] Step 2: For each candidate splitting feature, select the optimal splitting point according to the information gain ratio. The formula for calculating the information gain ratio is

[0111]

[0112] where is the information gain ratio of feature to the data set

[0113]

[0114] is the information gain

[0115]

[0116] is the information entropy of the data set is the number of categories in the data set, is the th data subset in the data set that belongs to the category, is the number of samples in the data set is the th sample subset in the data set where the value of feature is

[0117]

[0118] is the intrinsic value of feature is the th possible value of feature

[0119] Step 3: Split the data set according to the optimal splitting point to generate child nodes.

[0120] Step 4: Recursively repeat Step 2 and Step 3 for each child node until the stopping condition is met. The stopping conditions include that the node data belongs to the same category or reaches the maximum depth.​​​​

[0121] Step 5: Based on the generated decision tree, make decisions on the extracted features to obtain the interactive decision result.

[0122] It should be noted that this module uses an improved decision tree algorithm to make interactive decisions based on the extracted features. The decision tree algorithm mentioned here is a commonly used classification algorithm that predicts the category of unknown data by learning data features and decision rules. In this system, the decision tree is used to make the most appropriate interactive decision according to the features input by the user.

[0123] Specifically, the improved decision tree algorithm includes initializing the decision tree, taking all features as candidate splitting features, and selecting the optimal splitting point according to the information gain ratio. The information gain ratio is an important indicator to measure the classification ability of features for a data set, and it is calculated by comparing the information entropy before and after feature splitting. The calculation formula of the information gain ratio involves the information entropy of the data set and the intrinsic value of the feature, and these parameters can be set according to the specific features and category distribution of the data set. In actual operation, the algorithm will recursively repeat the steps of selecting the optimal splitting point and splitting the data set for each sub-node until the stopping condition is met, such as the node data belonging to the same category or reaching the maximum depth.

[0124] Preferably, the operation steps of the intelligent decision module can be further refined or alternative solutions can be provided. For example, when initializing the decision tree, different strategies can be adopted to select the initial feature set, such as the random forest method or sorting based on feature importance. When selecting the optimal splitting point, in addition to the information gain ratio, other indicators such as Gini impurity or mean squared error can also be considered to evaluate the splitting effect.

[0125] Furthermore, during the recursive splitting process, different stopping conditions can be set, such as the minimum sample number or the maximum depth limit, to prevent overfitting. In addition, regularization techniques can be introduced to control the model complexity, or ensemble learning methods such as gradient boosting trees can be used to improve the stability and accuracy of the decision tree.

[0126] In some embodiments, the content generation module includes the following steps:

[0127] Step 1: Determine the type of interactive content to be generated according to the decision result of the intelligent decision module. The types of interactive content include text and images.

[0128] Step 2: If it is of the text type, select a suitable template from the predefined text template library and combine the key information in the user input to generate specific text content through a text filling algorithm The text filling formula is where Represents a text filling function for filling key information into a template to generate specific text content .

[0129] It should be noted that this module is responsible for determining the type of interaction content to be generated based on the decision result of the intelligent decision-making module, and selecting a suitable template from a predefined template library to generate specific content. Here, the content generation module refers to a part of the system that creates user-understandable output based on the decision result, and the interaction content type refers to the form of the output, such as text or image.

[0130] Specifically, the content generation module first determines the type of interaction content based on the output of the intelligent decision-making module. If the decision result indicates that text-type interaction content needs to be generated, the module will select a suitable template from the predefined text template library. The text template library is a set of pre-designed text frameworks that can be customized according to different interaction scenarios and requirements. Then, combined with the key information in the user input, specific text content is generated through a text filling algorithm. The text filling algorithm is a method of inserting specific information into a template to form a complete sentence or paragraph.

[0131] Preferably, the operation steps of the content generation module can be further refined. For example, when determining the interaction content type, a series of rules can be set or a machine learning model can be used to automatically judge the most suitable content type. When selecting a template, the selection process of the template can be optimized based on the user's historical interaction data and current context information. In the text filling algorithm, natural language processing techniques, such as conditional random fields (CRF) or sequence-to-sequence models (Seq2Seq), can be adopted to more naturally integrate key information into the template.

[0132] Furthermore, machine learning techniques can also be considered to continuously optimize the template library to make it more in line with the actual needs and preferences of users. For the generation of image content, image processing and computer vision techniques, such as generative adversarial networks (GANs) or variational autoencoders (VAEs), can be integrated to create or modify image content.

[0133] In some embodiments, the feedback optimization module includes the following steps:

[0134] Step 1: Collect feedback information from users on the interaction content , and the feedback information includes the user satisfaction score , the user satisfaction score ranges from , and the text feedback content .

[0135] Step 2: According to the feedback information , if the user satisfaction score is lower than the set threshold, where the set threshold is , adjust the relevant decision tree model parameters. The adjustment formula is

[0136]

[0137] where is the adjusted parameter, is the parameter before adjustment, is the learning rate, is the gradient of the loss function calculated based on the feedback information with respect to the parameter .

[0138] It should be noted that this module involves collecting feedback information from users on the interaction content and optimizing the system performance based on this feedback. Here, the feedback information refers to the direct responses of users to the interaction content generated by the system, including satisfaction scores and text feedback content. The user satisfaction score is a quantitative indicator used to measure the satisfaction degree of users with the interaction content.

[0139] Specifically, the feedback optimization module first collects feedback information from users on the interaction content, which includes user satisfaction scores and text feedback content. The user satisfaction score is a rating system from 1 to 5, where 1 means very dissatisfied and 5 means very satisfied. The text feedback content can be specific comments or suggestions from users on the interaction content. Based on this feedback information, if the user satisfaction score is lower than the set threshold, such as 3, the system will adjust the relevant decision tree model parameters. The set threshold here is a preset standard used to determine when to adjust the system according to user feedback.

[0140] Preferably, the operation steps of the feedback optimization module can be further refined. For example, when collecting feedback information, an online survey or an automatically popped-up rating window can be used to obtain the user satisfaction score and text feedback. During the parameter adjustment process, the gradient descent algorithm can be used to update the parameters of the decision tree model, where the learning rate is a key hyperparameter that controls the step size of parameter update.

[0141] Furthermore, advanced machine learning techniques, such as reinforcement learning, can be introduced to automatically adjust the system parameters to better adapt to user feedback. A feedback analysis engine can also be set up, which can automatically identify keywords and sentiment tendencies in user feedback, thereby more accurately guiding system optimization. For the text feedback content, natural language processing techniques, such as sentiment analysis, can be applied to evaluate the emotional responses of users and adjust the system response strategy accordingly.

[0142] In some embodiments, during the training process of the intent classification neural network model, the cross-entropy loss function is adopted, and the formula of the loss function is

[0143]

[0144] where is the cross-entropy loss value, is the number of training samples, is the total number of intent categories, is the sample belongs to the intent category true label, if the sample belongs to the intent category then is otherwise it is , is the probability that the model predicts that the sample belongs to the intent category , is the sample corresponding word vector sequence.

[0145] It should be noted that this module involves adopting the cross-entropy loss function during the training process of the intent classification neural network model. Here, the cross-entropy loss function is a metric that measures the difference between the predicted probability distribution of the model and the true label probability distribution, and is commonly used in classification problems, especially during the training process of neural networks. It guides the learning of the model by calculating the difference between the predicted probability and the actual label.

[0146] Specifically, the calculation of the cross-entropy loss function involves the number of training samples, the total number of intent categories, the true label of the sample belonging to the intent category, and the probability that the model predicts the sample belongs to the intent category. In this process, the true label can be represented by one-hot encoding, that is, if the sample belongs to the intent category i, then the i-th element in its label vector is 1 and the rest are 0. The probability predicted by the model is the activation value of the output layer of the neural network, usually obtained through the softmax function. The specific calculation formula of the loss function involves summing over all samples and all categories, as well as logarithmic operations.

[0147] Preferably, the operation steps of training the intent classification neural network model can be further refined. For example, when setting the training parameters, appropriate learning rate and batch size can be selected, and these parameters will affect the convergence speed and stability of the model. The learning rate determines the amplitude of parameter update in each iteration, while the batch size determines the number of samples used for updating parameters each time.

[0148] Furthermore, regularization techniques such as L1 or L2 regularization can be adopted to prevent the model from overfitting. In terms of model structure selection, different types of neural network structures such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs) can be considered to adapt to the characteristics of different input data. For optimization algorithms, in addition to the commonly used stochastic gradient descent (SGD), more advanced optimizers such as Adam or RMSprop can also be considered to accelerate the model training process and improve performance.

[0149] In some embodiments, during the decision tree generation process, a pre-pruning strategy is adopted. When the information gain ratio of a node is less than a set threshold the splitting of this node is stopped.

[0150] It should be noted that this module involves the adoption of a pre-pruning strategy during the decision tree generation process. Here, the pre-pruning strategy is a technique used to prevent overfitting when constructing a decision tree, which is achieved by stopping the splitting before the tree is fully grown. This method helps to improve the generalization ability of the model and avoid reducing the performance of the model on new data due to overfitting the training data.

[0151] Specifically, the application of the pre-pruning strategy in the decision tree generation process includes stopping the splitting of a node when its information gain ratio is less than the set threshold. The information gain ratio is an important indicator for measuring the classification ability of a feature for a data set, which is calculated by comparing the information entropy before and after feature splitting. The set threshold is a pre-determined value used to control the growth of the decision tree and prevent the tree from becoming too complex. In actual operation, when the calculated information gain ratio is lower than this threshold, the algorithm will no longer perform further splitting on the current node.

[0152] Preferably, the operation steps of the pre-pruning strategy can be further refined. For example, when setting the threshold of the information gain ratio, the optimal threshold can be determined based on the performance on the validation set to balance the complexity and accuracy of the model. In addition, cross-validation can be introduced to more robustly evaluate the model performance under different threshold settings. During the construction of the decision tree, different splitting criteria such as Gini impurity or mean squared error can be considered to increase the robustness of the model.

[0153] Furthermore, for an alternative to pre-pruning, a post-pruning strategy can be considered, that is, first let the decision tree grow completely, and then prune from the bottom according to the performance metrics. Cost complexity pruning can also be considered, which is a method that combines the advantages of pre-pruning and post-pruning by introducing a penalty term to balance the complexity of the tree and the fitting degree of the training data.

[0154] The above-mentioned various embodiments of the present invention have the following beneficial effects: The artificial intelligence-based interaction system described in the present invention can provide a highly integrated solution to enhance the intelligence level of human-computer interaction. The system accurately analyzes the user input through the user intention recognition module, uses the historical data storage module to provide a data basis for algorithm analysis, the feature extraction module extracts key features from the user input, the intelligent decision-making module makes interaction decisions using an improved decision tree algorithm, the content generation module generates corresponding interaction content according to the decision result, and the feedback optimization module collects user feedback to optimize the system performance. This comprehensive method can ensure that the system is more efficient and accurate in understanding user needs and providing appropriate responses.

[0155] In addition, the design of the system allows the use of the cross-entropy loss function when training the user intention recognition module, and the intelligent decision-making module adopts a pre-pruning strategy during the decision tree generation process. The application of these technologies can improve the training efficiency of the model and the generalization ability of the decision tree. Through the application of these technologies, the system can not only reduce the risk of overfitting but also maintain high adaptability and accuracy when processing new user inputs, thus providing a more stable and reliable interaction experience in practical applications.

[0156] As Figure 2 shown, an artificial intelligence-based interaction method 200 in some embodiments, the method 200 includes:

[0157] Step 201, analyze the user input content and identify the user intention;

[0158] Step 202, extract key features from the user input and related data;

[0159] Step 203, use an improved decision tree algorithm to make interaction decisions based on the extracted features;

[0160] Step 204, generate corresponding interaction content according to the decision result of the intelligent decision-making module;

[0161] Step 205, collect user feedback on the interaction content and optimize the system performance.

[0162] It can be understood that the steps described in the artificial intelligence-based interaction method 200 correspond to the respective modules in the artificial intelligence-based interaction system described with reference to Figure 1 Therefore, the modules, features, and beneficial effects described above for the artificial intelligence-based interaction system also apply to the artificial intelligence-based interaction method 200 and the operations included therein, and will not be elaborated herein.

[0163] Next, refer to Figure 3, which shows a schematic structural diagram of an electronic device 300 suitable for implementing some embodiments of the present invention. The electronic devices in some embodiments of the present invention may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), PMPs (Portable Multimedia Players), in-vehicle terminals (such as in-vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 3 The terminal device shown is merely an example and should not impose any limitations on the functions and usage scope of the embodiments of the present invention.

[0164] As Figure 3 shown, the electronic device 300 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 301, which may perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 302 or the program loaded from the storage device 308 into the random access memory (RAM) 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 are also stored. The processing device 301, the ROM 302, and the RAM 303 are connected to each other through a bus 304. The input / output (I / O) interface 305 is also connected to the bus 304.

[0165] Generally, the following devices may be connected to the I / O interface 305: an input device 306 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 307 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 308 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 309. The communication device 309 may allow the electronic device 300 to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 3 the electronic device 300 with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. More or fewer devices may be alternatively implemented or had. Figure 3 Each block shown in

[0166] Further, the storage medium according to the embodiment of the present application stores program instructions capable of implementing all the above methods. Among them, the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: various media capable of storing program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.

[0167] The above description is only some preferred embodiments of the present invention and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. An interactive system based on artificial intelligence, characterized in that: include: User intention recognition module, used to parse user input content and identify user intention; The historical data storage module stores the historical data of past human-computer interactions; the historical data storage module stores the records of each interaction as follows: in, Indicates the number of interactions. For the User input for each interaction, For the The user intent identified by the interaction, For the Output of the sub-interactive system; The feature extraction module extracts key features from user input and related data for algorithm analysis; the feature extraction module includes the following steps: Step 1: Extract key features from user input text Extract word frequency features from the historical data storage module; Step 2: Calculate the user input With history input Similarity features of The intelligent decision module uses an improved decision tree algorithm to make interactive decisions based on the extracted features. The improved decision tree algorithm includes the following steps: Step 1: Initialize the decision tree and use all features as candidate split features; Step 2: For each candidate split feature, select the optimal split point based on the information gain rate; Step 3: According to the optimal split point, the data set is Split and generate child nodes; Step 4: recursively repeat step 2 to select the optimal split point and step 3 to generate child nodes for each child node until the stopping condition is met, which includes that the node data belongs to the same category or reaches the maximum depth; Step 5: Based on the generated decision tree, make decisions on the extracted features to obtain interactive decision results; The content generation module generates corresponding interactive content according to the decision results of the intelligent decision module; The feedback optimization module collects user feedback on the interactive content and optimizes system performance; wherein the user intention recognition module includes the following steps: Step 1: Text input by the user Perform word segmentation to obtain word sequence ,in, is the number of words, Representing word sequence The words; Step 2: Use the pre-trained word vector model to transform the word sequence Each word in Mapping to word vectors , get the word vector sequence ; Step 3: Sequence word vectors Input into the intent classification neural network model to calculate the intent categories The probability is: in, Indicates that in a given word vector sequence In the case of The probability of is the corresponding intent category output by the intent classification neural network model The score, is the total number of intent categories, and the intent category with the highest probability is finally determined as the identified user intent.

2. The artificial intelligence-based interactive system according to claim 1, characterized in that: In the feature extraction module, the user inputs text Extract word frequency features, and the word frequency calculation formula is: in, Expressive words When the user enters text The word frequency in Yes word In text The number of times it appears in is text The total number of times all words appear in ; Evaluating user input With history input The similarity feature of , the similarity calculation formula is: in, Represents user input With history input The similarity of and Separate words On user input and history input The frequency of words in .

3. The artificial intelligence-based interactive system according to claim 2, characterized in that: The optimal split point is selected according to the information gain rate. The information gain rate calculation formula is: in, It is a feature For the dataset The information gain rate of ； in, It is a dataset The information entropy of is the number of categories in the dataset, It is the first A subset of the data of the class, It is a dataset The number of samples, It is a dataset Medium Features The value is A subset of samples; in, It is a feature The intrinsic value of It is a feature All possible values ​​of .

4. The artificial intelligence-based interactive system according to claim 3, characterized in that: The content generation module comprises the following steps: Step 1: According to the decision result of the intelligent decision module, determine the type of interactive content to be generated, and the interactive content type includes text and image; Step 2: If it is a text type, select a suitable template from the predefined text template library , combined with key information from user input , generate specific text content through text filling algorithm , the text filling formula is ,in Represents a text filling function, which is used to fill in key information Fill in template Generate specific text content in .

5. The artificial intelligence-based interactive system according to claim 4, characterized in that: The feedback optimization module includes the following steps: Step 1: Collect user feedback on interactive content , feedback information includes user satisfaction ratings , user satisfaction rating The value range is , and text feedback content ; Step 2: Based on feedback information , if the user satisfaction score Below the set threshold, the set threshold is , adjust the relevant decision tree model parameters, and the adjustment formula is: in, is the adjusted parameter, is the parameter before adjustment, is the learning rate, Based on feedback information The calculated loss function is the parameter gradient.

6. The artificial intelligence-based interactive system according to claim 5, characterized in that: In the training process of the intent classification neural network model, the cross entropy loss function is used, and the loss function formula is: in, is the cross entropy loss value, is the number of training samples, is the total number of intent categories, It is a sample Belongs to the intent category The true label of the sample Belongs to the intent category but for , otherwise , is the model prediction sample Belongs to the intent category The probability of It is a sample The corresponding word vector sequence.

7. The artificial intelligence-based interactive system according to claim 6, characterized in that: In the decision tree generation process, the pre-pruning strategy is adopted. When the information gain rate of the node Less than the set threshold When , stop splitting the node.

8. An artificial intelligence-based interactive method, applied to the artificial intelligence-based interactive system according to any one of claims 1 to 7, characterized in that: include: Parse user input and identify user intent; Extract key features from user input and related data; An improved decision tree algorithm is used to make interactive decisions based on the extracted features; Generate corresponding interactive content based on the decision results of the intelligent decision-making module; Collect user feedback on interactive content and optimize system performance.

Images