WO2020147428A1 - 交互内容生成方法、装置、计算机设备及存储介质 - Google Patents

交互内容生成方法、装置、计算机设备及存储介质 Download PDF

Info

Publication number
WO2020147428A1
WO2020147428A1 PCT/CN2019/120595 CN2019120595W WO2020147428A1 WO 2020147428 A1 WO2020147428 A1 WO 2020147428A1 CN 2019120595 W CN2019120595 W CN 2019120595W WO 2020147428 A1 WO2020147428 A1 WO 2020147428A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
parameter
current
intention
information
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/120595
Other languages
English (en)
French (fr)
Inventor
柳明辉
徐国强
邱寒
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
OneConnect Smart Technology Co Ltd
Original Assignee
OneConnect Smart Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by OneConnect Smart Technology Co Ltd filed Critical OneConnect Smart Technology Co Ltd
Publication of WO2020147428A1 publication Critical patent/WO2020147428A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/335—Filtering based on additional data, e.g. user or group profiles
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification

Definitions

  • This application relates to the technical field of semantic analysis, and in particular to a method, device, computer equipment and storage medium for generating interactive content.
  • chatbots focus on a single round of interaction with the user, so that the robot cannot obtain the true intention of the user well, because the important information of the user chat may be in the user's previous conversation.
  • This single-round-based chat mechanism ignores the current user’s previous rounds of chat topics and scene analysis, so the response content returned may be biased or even wrong. How to obtain chat content closer to the user's true intentions has become an urgent problem to be solved.
  • the embodiments of the present application provide a method, device, computer device, and storage medium for generating interactive content to solve the problem of how to obtain chat content that is closer to the user's real intention.
  • a method for generating interactive content including:
  • a preset reinforcement learning model to analyze at least one current intention parameter and at least one historical intention parameter to obtain a target intention, the target intention including at least one target parameter and parameter priority order;
  • the information conversion model is called to convert each retrieved text, obtain the current reply information corresponding to each retrieved text, and push at least one current reply information to the client according to the parameter priority order.
  • An interactive content generating device including:
  • the current information receiving module is configured to receive the current round information carrying the session identifier sent by the client, and obtain at least one current intention parameter based on the current round information;
  • a historical parameter obtaining module configured to query the session record database based on the session identifier, and obtain at least one historical intention parameter corresponding to the session identifier;
  • An acquiring target intention module configured to use a preset reinforcement learning model to analyze at least one current intention parameter and at least one historical intention parameter to acquire a target intention, the target intention including at least one target parameter and parameter priority order;
  • the retrieval text acquisition module is used to query and retrieve the text database based on each target parameter, and obtain the retrieval text corresponding to each target parameter;
  • the reply information obtaining module is used to call the information conversion model to convert each retrieved text, obtain the current reply information corresponding to each retrieved text, and push at least one current reply information to the client according to the parameter priority order.
  • a computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor.
  • the processor executes the computer-readable instructions, the following steps are implemented:
  • One or more readable storage media storing computer readable instructions
  • the computer readable storage medium storing computer readable instructions
  • the one Or multiple processors perform the following steps:
  • FIG. 1 is a schematic diagram of an application environment of a method for generating interactive content in an embodiment of the present application
  • Figure 2 is a flowchart of a method for generating interactive content in an embodiment of the present application
  • FIG. 3 is a schematic diagram of the realization process of obtaining the goal intention in an embodiment of the present application.
  • FIG. 4 is another flowchart of a method for generating interactive content in an embodiment of the present application.
  • FIG. 5 is another flowchart of a method for generating interactive content in an embodiment of the present application.
  • Fig. 6 is another flowchart of a method for generating interactive content in an embodiment of the present application.
  • FIG. 7 is another flowchart of a method for generating interactive content in an embodiment of the present application.
  • Fig. 8 is a schematic diagram of an interactive content generating apparatus in an embodiment of the present application.
  • Fig. 9 is a schematic diagram of a computer device in an embodiment of the present application.
  • the interactive content generation method provided by the embodiments of the present application can be applied in an application environment as shown in FIG. 1.
  • the interactive content generation method is applied in an interactive content generation system.
  • the interactive content generation system includes a client and a server. Communicate with the server through the network.
  • the client is also called the client, which refers to the program that corresponds to the server and provides local services to the client.
  • the client can be installed on, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices and other computer devices.
  • the server can be implemented by an independent server or a server cluster composed of multiple servers, which is used to accept the current round information sent by the user through the client and generate the reply information corresponding to the chat robot.
  • a method for generating interactive content is provided.
  • the method is applied to the server in FIG. 1 as an example for description, including the following steps:
  • the current round information is the information that the user enters into the client in the current round to express the user's intention. Further, the current round information may include intention information expressed in various forms, including but not limited to text data, voice information, or gesture action information, which is not limited here.
  • the session identifier is an identifier used to distinguish different session scenarios initiated by the server. Since the server can start several session scenarios with several clients, in order to distinguish each session scenario for session content analysis, the server needs to set a corresponding session identifier for each session scenario.
  • Intent is the purpose expressed by the user in the current round of information. It can be divided into “explicit intent” and “implicit intent”. The difference between the two is very obvious:
  • the so-called explicit intention means that a type of vocabulary expressing intention clearly appears in the user's words, such as: “hope”, “want”, “need” and so on.
  • this kind of display intention judgment is less difficult. It only needs to identify these fixed intention words, and then make connections with other components in the sentence. For example, if the user enters "I want to book a ticket to Beijing", the chat robot can recognize the intention word "want”, and get the intention parameters of the sentence to book the air ticket for me.
  • the air ticket itinerary includes the current location to Beijing.
  • Implicit intent is the opposite of explicit intent. There is no vocabulary that directly reflects the intent in the user's utterance. The chat robot needs to judge the user's intent based on text data. For chat bots, implicit intentions are more difficult to judge.
  • One of the more commonly used methods is to convert implicit intentions into explicit intentions before processing. For example, when a user expresses the implicit intention of "I am hungry", it is first transformed into the corresponding explicit intention form of "I want to eat”, and then the chatbot is processed according to the explicit intention.
  • the current intention parameter is the verbs and nouns (de-stop words) with actual meaning extracted from the text data after converting the current round information into corresponding text data, and express the user's intention in the most concise way.
  • stop words mainly include English characters, numbers, mathematical characters, punctuation marks, and function words with high frequency of use.
  • the text data is "Today's weather is really good”
  • the current intention parameters extracted from the text data include: today, weather and good (excluding the stop words "true” and "ah”).
  • step S10 the server can extract at least one current intention parameter through the current round information sent by the client, filter out the meaningless function words in the current round information, and directly obtain valid current intention parameters in the current round information , To prepare the technical basis for the subsequent combination of historical intention parameters to obtain the user’s true intention.
  • the session record database is a set of session records saved by the server based on each session scene (that is, each session identifier).
  • the historical intent parameters correspond to the current intent parameters, and are based on the intent parameters of all rounds saved before the current round of the session identified by the same session.
  • the server correspondingly records the current intention parameters generated by the current round of conversations in the conversation record database based on the same conversation identifier. Understandably, each current intent parameter is stored in the session record database to form a historical intent parameter, so that the subsequent server can match all corresponding historical intent parameters based on the same session identifier for real intent analysis.
  • step S20 the server can directly obtain all corresponding historical intent parameters in the session record database based on the session identifier, without the server re-extracting historical intent parameters in all historical sessions, and accelerate the processing speed of the server in analyzing the user's real intent.
  • the reinforcement learning model is preset on the server to analyze all historical intention parameters and current intention parameters input to the model, and obtain the analysis result as the target intention model.
  • the server may use a combination of an LSTM (Long Short-Term Memory) model and a softmax classifier as the reinforcement learning model.
  • LSTM Long Short-Term Memory
  • multi-level target intentions can be set for the chat scene.
  • the first-level target intention includes chat and questioning; the first-level target intention can continue to be divided into multiple second-level target intentions, and the chat intentions in the first-level target intention can be Continue to be divided into life, work and leisure; based on the second-level goal intention, it can be further divided into the third-level goal intention, and continue to be refined according to the needs of the scene.
  • the implementation process for the server to obtain the target intention through text data is as follows:
  • Preprocess the text data including removing punctuation marks from the corpus, removing stop words (words with no actual meaning), etc.
  • the word2vec tool is a NLP (Natural Language Processing, natural language processing) tool, which can vectorize all natural language words into a dense vector that the computer can understand (Dense Vector) for quantitative measurement The relationship between words, mining the relationship between words. Understandably, for similar words, their corresponding word vectors are also similar.
  • NLP Natural Language Processing, natural language processing
  • the LSTM model can solve the problem that the natural language processing statistical method can only consider the recent n words and ignore the words longer ago. It aims to find the correlation between words and words, increase the time text content in data analysis, and remember What happened before is then applied to the neural network to observe the connection with what happens next to the neural network to get the goal intention.
  • LSTM The characteristic of LSTM is that in addition to the RNN (Recurrent Neural Network) model, valve nodes of each layer are added, as shown in Figure 4. There are three types of valve nodes: forget gate, input gate and output gate. These valve nodes can be opened or closed, and are used to determine whether the memory state of the RNN model (the state of the previous network) in the previous output layer has reached the threshold, and thus is added to the calculation of the current layer.
  • RNN Recurrent Neural Network
  • the valve node uses the sigmoid function to calculate the calculation result of the previous output layer corresponding to the valve node in the RNN model; if the calculation result reaches the threshold, multiply the valve node output and the calculation result of the current layer as the input of the next layer ; If the threshold is not reached, forget the calculation result of the previous output layer corresponding to the valve node.
  • the memory function of the LSTM model is realized by these valve nodes.
  • the training results of the previous RNN model will be associated with the current RNN model for calculation, and when the valve node is closed, the previous calculation results will no longer affect the current calculation. Therefore, by adjusting the switch of the valve node, it is possible to realize the influence of the historical intention on the final acquisition of the target intention.
  • future intent analysis such as starting to analyze new paragraphs or chapters in natural language processing, just turn off the valve node.
  • the softmax classifier is introduced into the neural network to define a new output layer (target intention probability distribution).
  • the output layer of the LSTM model based on the neural network is not necessarily a probability distribution layer, so the softmax classifier can be introduced to the LSTM model at the end.
  • the softmax classifier can be used as an additional processing layer to take the output of the neural network (the final output of the LSTM model) Layer) becomes a probability distribution, that is, each output is a decimal between 0 and 1, and the sum of all output results is 1. For example, if the final output layer of the LSTM model is y1, y2, y3 and y4, the server can use the softmax classifier to process the final output layer of the LSTM model, using the following formula:
  • the value of y′ 1 y′ 2 y′ 3 y′ 4 can be calculated by the softmax classifier, where y′ i is the output probability of each target intention. Understandably, the target intention corresponding to the output with the largest value in y′ i is the target intention corresponding to the current round information.
  • the text data "I recently watched the new movie directed by Jiang Wen to suppress the evil”
  • the target intention obtained after recognition by the preset intention recognition model in step S20 is: the user discusses the movie, and the movie parameters Including: The movie is called "Xie Bu Zheng Zheng".
  • this embodiment can also set the parameter priority order for the target parameter, that is, set a higher parameter priority order for the noun corresponding to the specified action object after the verb, and set the secondary parameter priority order for the action executor .
  • the focus of people talking about events is the object of action execution, that is, the priority order of parameters of the object of action execution is generally higher than that of the person who performs the action.
  • Setting parameter priority order for different target parameters facilitates subsequent conceptual expansion based on different parameter priority order.
  • the server may also set the priority value of the parameter including the target intention with the largest number of target parameters to also be the highest.
  • the text data entered by the user through the client is "Query Longgang rental housing information, with an area of 20 square meters and a cheap price”.
  • the server can analyze that the target intention of the text data is to rent a house in Longgang, and split the text data for analysis:
  • the third group includes the most target parameters, that is, the parameter priority order of the third group is 3, and so on, the parameter priority order of the second group is 2, and the parameter priority order of the first group The order is 1.
  • the server responds based on the parameter priority order, it can search in descending order of the parameter priority order, that is, the third group of queried house rental results are first returned to the client.
  • the server obtains the realization process of the target parameter through the target intention:
  • the target parameters of the template can be obtained as "action person” and "leisure activity”. Extract the words corresponding to the above target parameters from the text data "I recently watched the new movie directed by Jiang Wen”: the action person corresponds to "I” (that is, the user), and the leisure activity "the movie does not suppress the evil” .
  • the parameter priority order corresponding to the actor is level 1
  • the parameter priority order corresponding to leisure activities is level 2.
  • step S30 the server can recognize the user's target intention through the preset reinforcement learning model, accurately determine the user's needs, and respond with the highest priority based on the parameter priority order corresponding to the target parameter, which will help to obtain the corresponding response based on the target parameter.
  • the reply message of the user accurately promotes the chat around the user’s focus to maintain the stickiness and practicality of the chat between the chat robot and the user.
  • the online database is an online search database, such as Baidu, Sogou, Google, or 360 search.
  • the search text is the reply text corresponding to the target parameter. For example, if the target parameter is "Today's Weather", the server can query the online weather through the network database to obtain the weather query result corresponding to "Today's Weather".
  • the weather query result is the search text.
  • the server may obtain the retrieval text corresponding to the target parameter through the online database, and provide the user with text information related to the current round information, thereby improving the accuracy of the reply content.
  • the information conversion model in this step is a model that converts text information into a form of reply information delivered to the user. For example, if the user inputs the current round information in the form of voice, the voice conversion model is used in this step to convert the retrieved text into voice information and output to the user, or it can be directly output to the client in text form according to the client's preferences.
  • the information conversion model is a model that mutually converts the meaning of text data and specific expressions, such as converting voice information into corresponding text data, and converting text data into corresponding sign language actions.
  • the information conversion model can be converted based on multiple expressions, so it includes multiple conversion models.
  • a voice conversion model that converts voice information into text information, such as RNN-HMM (Recurrent Neural Network-Hidden Markov Model, or Recurrent Neural Network-Hidden Markov) model or LSTM-HMM (Long Short-Term Memory), long and short term Memory network-Hidden Markov) model and other machine acoustic models, or gesture conversion models that convert gesture information into text information, such as FLDCRFs (Fuzzy based Latent dynamic Condition Random Fields) and other machine gesture recognition models .
  • RNN-HMM Recurrent Neural Network-Hidden Markov Model, or Recurrent Neural Network-Hidden Markov
  • LSTM-HMM Long Short-Term Memory
  • Memory network-Hidden Markov model Long Short-Term Memory
  • gesture conversion models that convert gesture information into text information, such as FLDCRFs (Fuzzy based Latent dynamic Condition Random Fields) and other machine gesture recognition models.
  • FLDCRFs Fuzzy based Latent dynamic Condition Random Fields
  • machine learning models such as the RNN-HMM machine acoustic model, the LSTM-HMM machine acoustic model, and the FLDCRFs machine gesture recognition model in this embodiment are well-known technologies, and will not be repeated here.
  • step S10 if the user inputs the current round information (voice information) through the microphone of the client terminal: "I recently watched the new movie by Jiang Wen Press positive".
  • the server obtains the current round information in the voice format collected by the microphone through the client, so it can determine that the current round information is voice information, and the RNN-HMM machine acoustic model can be used to convert the voice information into "I recently watched director Jiang Wen’s The text data of "new movie evil does not suppress the right”.
  • step S60 in order to maintain the consistency of the way of communicating with the user and improve the fun of the chat, the chat computer can also use the RNN-HMM machine acoustic model for the retrieved text obtained in step S50, and convert the retrieved text into voice information (current reply information) Return to the client.
  • the server may return the current reply information corresponding to the retrieved text to the client according to the priority order of the parameters, which is conducive to direct reply based on the user's focus, and promptly returns accurate current reply information for reference to the client to improve
  • the relevance of the chat content between the chat bot and the user the server can be set according to the needs of the scene so that the form of the output reply information is consistent with the way the user enters the current round of information, maintaining the consistency of chat habits, and improving the user’s and the chat Chatting is fun.
  • the server analyzes and obtains at least one current intention parameter by receiving the current round information sent by the client, and combines the current intention parameter with the session record database to obtain at least one historical intention parameter to obtain the target intention and Corresponding reply information can ensure that the chat bot generates accurate reply information to the client in a timely manner, avoiding inaccurate or irrelevant reply information only relying on the current intention parameters, and improving the interaction and reliability between the chat bot and the client .
  • receiving the current round information sent by the client includes:
  • the text data is the text information corresponding to the meaning of the current round information sent in a specific way.
  • the user sends a nodding action through the camera of the client as the current round information
  • the server can call the information conversion model (in this embodiment, the action The recognition model) converts the nodding action into text data "agree”.
  • step S11 the server converts the received current round information sent by the client into text data, which facilitates the server to perform further processing based on the text data to obtain the reply information corresponding to the current round information and return it to the client.
  • S12 Use the preset language processing model to analyze and process the text data to obtain at least one current intention parameter.
  • the preset language processing model of this embodiment can use word2vec, a NLP (Natural Language Processing, natural language processing) tool, which can vectorize all natural language words into dense vectors that can be understood by computers (Dense Vector), which is used to quantitatively measure the relationship between words and explore the relationship between words. Understandably, for similar words, their corresponding word vectors are also similar.
  • NLP Natural Language Processing, natural language processing
  • step S10 a preset language processing model is used to analyze and process the text data, and the realization process of obtaining at least one current intention parameter has been described in detail in step S10. In order to avoid repetition, it will not be repeated here.
  • the server may use a preset language processing model to parse the text data to obtain at least one current intention parameter, and may extract the most concise current intention expressed by the user in the current round, which is beneficial for the subsequent server to combine all historical intention parameters based on the current intention Get the real intention of the user.
  • the server converts the received current round information sent by the client into text data, which facilitates the server to further process the text data based on subsequent processing to obtain the reply information corresponding to the current round information and return it to the client end.
  • the server can use a preset language processing model to parse the text data to obtain at least one current intention parameter, and can extract the most concise current intention expressed by the user in the current round, which is conducive to the subsequent server to obtain the user’s truth based on the current intention and all historical intention parameters. intention.
  • step S30 using a preset reinforcement learning model to analyze at least one current intention parameter and at least one historical intention parameter to obtain the target intention, including:
  • the word vector is actually the semantic mapping matrix obtained by mapping the word to a semantic space.
  • the central word A maps the peripheral words BCDAEFG to obtain a parameter matrix W1
  • the central word L is mapped to the peripheral words BCDLEFG to obtain a parameter matrix W2. If the distance between W2 and W1 is similar, it means that A and L can be mapped to the same surrounding words, which may be synonyms.
  • Each word has a one-hot vector whose dimension is V. Among them, the one-hot vector is used to characterize that each element in the vector is associated with a word in the lexicon.
  • the vector of the specified word is expressed as: its corresponding element in the vector is set to 1, and the other elements are set to 0 (if The word has appeared in the vocabulary, then the corresponding position in the vocabulary in the vector is 1, and the other positions are all 0). If it does not appear in the vocabulary, the vector is all zeros.
  • the CBOW model continuous bag-of-words model, which takes the context of the word as input to predict the word itself
  • the implementation steps are as follows:
  • the size of batch_size must be an integer multiple of 2*window to ensure that each batch contains all samples corresponding to a word.
  • the neural network is iteratively trained for a certain number of times to obtain a parameter matrix from the input layer to the hidden layer with dimension N.
  • the transposition of each row in the matrix is the word vector matrix with the dimension V of the corresponding word.
  • the server may obtain the current word vector matrix and the historical word vector matrix corresponding to the current intent parameter and the historical intent parameter, respectively, to prepare a technical basis for the subsequent server to obtain the target intent based on the word vector matrix.
  • the combination of the LSTM (Long Short-Term Memory) model and the softmax classifier in step S30 can be used as the reinforcement learning model to analyze the current word vector matrix and the historical word vector matrix Analyze and process to obtain target intent.
  • LSTM Long Short-Term Memory
  • the current word vector matrix and the historical word vector matrix are input to the input valve of the preset reinforcement learning model.
  • the specific implementation process has been described in detail in the foregoing step S30. In order to avoid repetition, it will not be repeated here.
  • step S32 the server can recognize the user's target intention through the preset reinforcement learning model, accurately determine the user's needs, obtain chat content that is more suitable for the user's needs, and reduce the number of unanswered questions between the chat robot and the user.
  • the server may obtain the current word vector matrix and the historical word vector matrix corresponding to the current intent parameter and the historical intent parameter, respectively, to prepare a technical basis for the subsequent server to obtain the target intent based on the word vector matrix.
  • the server can identify the user's target intention through the preset reinforcement learning model, accurately determine the user's needs, obtain chat content that is more suitable for the user's needs, and reduce the number of unanswered questions between the chatbot and the user.
  • step 40 querying a networked database based on each target parameter to obtain the retrieval text corresponding to each target parameter specifically includes the following steps:
  • the attribute of the target parameter is a product attribute
  • the product attribute is a thing attribute marked by the content provider for each target parameter through the server.
  • the content provider is a provider that provides the server where the chat robot is located, and the provider can sell other physical products or provide various services at the same time.
  • the chatbot can recommend marketing products or services provided by the content provider at the right time when the user communicates, which is beneficial to the improvement of content providers using the chatbot system.
  • Product or service sales conversion rate For example, if the target parameter is "insurance”, if the insurance is a free product of the content provider, the content provider can mark "insurance" as a product attribute.
  • the server can associate the physical name or service name in each target intention template with the physical object or service type provided by the content provider itself, so that when the target parameter obtained by the server includes the physical name or service name.
  • Content providers can introduce corresponding products or services to users based on the association relationship.
  • the product database is a database that stores all the products or services that the content provider can provide.
  • Product recommendation information is an introduction to the product content of each product or service record in the product database.
  • the product recommendation information may include a product summary introduction and a detailed product introduction, so that the server first pushes the product summary introduction to the client.
  • the server receives the detailed product introduction request sent by the client, it indicates that the user is willing to continue to understand the details of the product, and then continue to send the detailed product introduction corresponding to the product to the client.
  • the server can pre-mark each target parameter with a label whether it belongs to the product attribute, which is helpful for the server to determine whether the target parameter is a product attribute immediately after obtaining the target parameter. time.
  • the server can obtain the product recommendation information corresponding to the target parameter based on the product database, and then push it to the user subsequently to improve the actual applicability of the chat robot and the scalability of chat content.
  • the interactive content generation method further includes: pushing the product recommendation information to the client in a preset product push format.
  • the preset product push format is a format preset by the server to push products to the client in the chat interface.
  • the server can set a list of product names to push, or a brief introduction to each product corresponding to the product.
  • the server when the server receives the product name selected by the client, it can associate the product name with the detailed product introduction established in step S51, so as to push the detailed product introduction to the client, so as to prevent users who are pushed to the client for the first time from feeling uncomfortable Detailed introduction of the product of interest.
  • the server can send product recommendation information to the client according to a preset product push format, so as to maintain the consistency of the format when the chat content involves products or services.
  • the online database is an online search database, such as Baidu, Sogou, Google, or 360 search.
  • the search text is the reply text corresponding to the target parameter. For example, if the target parameter is "Today's Weather", the server can query the online weather through the network database to obtain the weather query result corresponding to "Today's Weather".
  • the weather query result is the search text.
  • step S42 the server can obtain the retrieval text corresponding to the target parameter through the online database, and provide the user with the retrieval text related to the current round information, thereby improving the accuracy of the reply content; at the same time, the server does not need to store the retrieval corresponding to the target parameter locally Text can save the local storage space of the server.
  • the server can pre-mark whether each target parameter belongs to the product attribute label, which is helpful for the server to determine whether the target parameter is a product attribute, and it is not necessary to analyze whether each target parameter is a product attribute, which is beneficial to shorten the server The judgment time.
  • the server can obtain the product recommendation information corresponding to the target parameter based on the product database, and then push it to the user subsequently to improve the actual applicability of the chat robot and the scalability of chat content.
  • the server can obtain the retrieval text corresponding to the target parameter through the networked database, and provide the user with the retrieval text related to the current round information to improve the accuracy of the reply content; at the same time, the server does not need to store the retrieval text corresponding to the target parameter locally, which can save Local storage space of the server.
  • the interactive content generation further includes:
  • the user's personal information is static data and dynamic data collected by the server.
  • the static data is the personal information that the user actively enters when registering on the server and will not change for a long time, such as the user's region, age, gender, culture, Occupation and income, etc.
  • dynamic data is the data obtained by analyzing user behaviors during the server's interaction with users, such as living habits or consumption habits.
  • the following user personal information is needed:
  • Demographic attributes including basic information such as gender and age;
  • Equipment attributes terminal characteristics used, etc.
  • Behavior data user's behavior log data on the website such as access time and browsing path;
  • the user portrait is a highly refined label abstracted by the server based on the user's personal information, that is, the user characteristic identification.
  • tags By "sticking" tags to users, users can be described with highly general and easy-to-understand features, which facilitates further processing by the server based on tags (standardized information).
  • the implementation process of generating a user portrait based on the user's personal information is as follows:
  • Tag Value Exhaustive tag value
  • User portraits include tags such as gender, age, mobile phone brand, place of residence, and hobbies.
  • Xiao Ming is an example of a user portrait.
  • the output results of Xiao Ming's user portrait are: "Male”, “18-35”, “iPhone”, “Beijing” and "Football”.
  • the server may construct a user portrait based on the user's personal information, and prepare a technical basis for the subsequent server to analyze the user's personality based on the user portrait.
  • the user's personality is a person's stable attitude towards reality, and the personality characteristics shown in the habituated behavior pattern corresponding to this attitude.
  • Personality is relatively stable once it is formed, but it is not static, but plastic.
  • Personality is different from temperament and more embodies the social attributes of personality.
  • the core of personality differences between individuals is personality differences.
  • the shopping tags in the user portraits include the average browsing time and the average number of comparisons of the same type.
  • the online shopping tag records of three users are as follows:
  • the average browsing time is less than 10 minutes, and the average number of comparisons is less than 5 times.
  • the average browsing time is between 10 and 20 minutes, and the average number of comparisons is between 5 and 20.
  • Hesitant The average browsing time is greater than 20 minutes, and the average number of comparisons is greater than 20.
  • the server compares the shopping tags of the three users with the shopping personality reference template to get the following results:
  • User B always compares a small number of similar products (average number of comparisons is between 5 to 20 times) within an appropriate amount of time (average browsing time is between 10 minutes and 20 minutes) before placing an order, then user B's shopping personality is rational.
  • the server can also analyze the user portrait to obtain the user's chat personality.
  • the chat personality includes: excited, calm, or low. It needs to be added that the server can also determine the user's personality by analyzing the voice characteristics in the current round information input by the user in combination with the chat personality reference template.
  • the chat personality reference template for voice characteristics is as follows:
  • Hearty voice If the user is male, he has a capable personality and strong self-esteem; if the user is a female, he has a lively personality, good social skills, strong self-esteem, and lacks a sense of humor.
  • a mellow and loud voice This group of people is full of energy, regardless of whether the user is male or female.
  • the server may analyze the user portrait obtained in step S601 based on actual experience, obtain the user personality corresponding to the user, and prepare a technical basis for the subsequent server to adopt different reply modes based on different user characteristics.
  • the reply pattern database is a database that is pre-installed on the server and is composed of reply information patterns for different user personalities.
  • the target reply mode is a chat mode that responds to the user's personality. This embodiment is mainly applied to voice chat, so the reply mode is the tone reply mode when the chat robot chats with the user.
  • the two parties whose voices are synchronized during chat are more likely to be accepted by the other party. That is to say, if the user speaks slowly, and if the chat robot chats with the user's rhythm, it can increase the user's willingness to continue chatting with the chat robot. Based on this, the server can match the corresponding chat tone according to the user's personality to further enhance the chatting fun and emotional interaction between the user and the chat robot.
  • the server can configure a tone response mode for each user's personality, for example, the tone response mode for user personality such as enthusiasm, excitement, or positivity is a quick and enthusiastic tone response mode; for calm, plain or mature users
  • the tone response mode of the personality is a medium-speed and calm tone response mode.
  • the server may match the corresponding target reply pattern in the preset reply pattern library according to the user's personality, so as to further enhance the chat interest and emotional interaction between the user and the chat robot.
  • the server sends the target reply mode (may be a tone reply mode) obtained in step S603, such as a quick and enthusiastic tone reply mode, etc., to the information conversion model, so that the information conversion model converts the conceptual text into a corresponding target reply mode. Matching current reply message. Further, the information conversion model can obtain the current reply information that matches the target reply mode by adjusting the speech rate and the tone of the speech.
  • the target reply mode may be a tone reply mode obtained in step S603, such as a quick and enthusiastic tone reply mode, etc.
  • the server may combine the target reply mode and the information conversion model to obtain an updated information conversion model to adjust the tone that matches the target reply mode, and send the corresponding reply information to the client to enhance the communication between the chat robot and the user Emotional interactivity to improve the effectiveness of chat content.
  • the server can analyze the user portrait based on actual experience to obtain the user's corresponding personality, and match the corresponding target reply pattern in the preset reply pattern library according to the user's personality to further enhance the relationship between the user and the chat robot
  • the chat is interesting and emotional.
  • the server can send the adjusted current reply information that matches the target reply pattern to the client, enhancing the emotional interaction between the chat robot and the user, and improving the effectiveness of the chat content.
  • the server analyzes and obtains at least one current intention parameter by receiving the current round information sent by the client, and combines the current intention parameter with the session record database to obtain at least one historical intention parameter to obtain the target intention and Corresponding reply information can ensure that the chat bot generates accurate reply information to the client in a timely manner, avoiding inaccurate or irrelevant reply information only relying on the current intention parameters, and improving the interaction and reliability between the chat bot and the client .
  • the server converts the received current round information sent by the client into text data, which facilitates the server to perform further processing based on the text data to obtain the reply information corresponding to the current round information and return it to the client.
  • the server can use a preset language processing model to parse the text data to obtain at least one current intention parameter, and can extract the most concise current intention expressed by the user in the current round, which is conducive to the subsequent server to obtain the user’s truth based on the current intention and all historical intention parameters. intention.
  • the server may obtain the current word vector matrix and the historical word vector matrix corresponding to the current intent parameter and the historical intent parameter, respectively, to prepare a technical basis for the subsequent server to obtain the target intent based on the word vector matrix.
  • the server can identify the user's target intention through the preset reinforcement learning model, accurately determine the user's needs, obtain chat content that is more suitable for the user's needs, and reduce the number of unanswered questions between the chatbot and the user.
  • the server can pre-mark each target parameter with a label as to whether it belongs to the product attribute, so that the server can determine whether the target parameter is a product attribute immediately after obtaining the target parameter, and it is not necessary to analyze whether each target parameter is a product attribute, which helps shorten the server's determination time .
  • the server can obtain the product recommendation information corresponding to the target parameter based on the product database, and then push it to the user subsequently to improve the actual applicability of the chat robot and the scalability of chat content.
  • the server can obtain the retrieval text corresponding to the target parameter through the networked database, and provide the user with the retrieval text related to the current round information to improve the accuracy of the reply content; at the same time, the server does not need to store the retrieval text corresponding to the target parameter locally, which can save Local storage space of the server.
  • the server can analyze the user portrait based on actual experience to obtain the user personality corresponding to the user, and match the corresponding target reply pattern in the preset reply pattern library according to the user personality to further enhance the chat between the user and the chat robot Interesting and emotional interaction.
  • the server can send the adjusted current reply information that matches the target reply pattern to the client, enhancing the emotional interaction between the chat robot and the user, and improving the effectiveness of the chat content.
  • an interactive content generating device is provided, and the interactive content generating device corresponds to the interactive content generating method in the foregoing embodiment one-to-one.
  • the interactive content generating device includes a current information receiving module 10, a historical parameter obtaining module 20, a target intent obtaining module 30, a target intent obtaining module 30, a retrieval text obtaining module 40 and a reply information obtaining module 50.
  • the detailed description of each functional module is as follows:
  • the current information receiving module 10 is configured to receive the current round information carrying the session identifier sent by the client, and obtain at least one current intention parameter based on the current round information.
  • the historical parameter obtaining module 20 is configured to query the session record database based on the session identifier, and obtain at least one historical intention parameter corresponding to the session identifier.
  • the target intention acquisition module 30 is configured to analyze at least one current intention parameter and at least one historical intention parameter by using a preset reinforcement learning model to acquire a target intention.
  • the target intention includes at least one target parameter and parameter priority order.
  • the retrieval text obtaining module 40 is used for querying and retrieving the text database based on each target parameter, and obtaining the retrieval text corresponding to each target parameter.
  • the reply information obtaining module 50 is used to call the information conversion model to convert each retrieved text, obtain the current reply information corresponding to each retrieved text, and push at least one current reply information to the client according to the parameter priority order.
  • the receiving current information module 10 includes a round information receiving unit 11 and an intention parameter obtaining unit 12.
  • the round information receiving unit 11 is configured to receive the current round information sent by the client, call the information conversion model to identify the current round information, and obtain text data.
  • the intent parameter obtaining unit 12 is configured to analyze and process text data using a preset language processing model to obtain at least one current intent parameter.
  • the module for obtaining the target intention includes obtaining a current word vector matrix and a historical word vector matrix unit, obtaining a part of speech vector matrix and obtaining a target intention unit respectively.
  • the current word vector matrix and the historical word vector matrix unit are respectively obtained, and the current word vector matrix and the historical word vector matrix are respectively obtained based on at least one current intention parameter and at least one historical intention parameter.
  • part-of-speech vector matrix which is used for part-of-speech tagging based on current intent parameters and historical intent parameters, and obtain the current part-of-speech vector matrix and historical part-of-speech vector matrix corresponding to the current word vector matrix and the historical word vector matrix respectively.
  • the target intention unit is used to analyze the current word vector matrix and the historical word vector matrix using a preset reinforcement learning model to obtain the target intention.
  • the acquiring retrieval text module includes an acquiring product information unit and an acquiring retrieval text unit.
  • the acquiring product information unit is used for, if the attribute of the target parameter is a product attribute, query the product database corresponding to the product attribute based on each target parameter, and obtain the product recommendation information corresponding to each target parameter as the retrieval text.
  • the retrieval text unit is used to query the online database based on each target parameter if the attribute of the target parameter is not a product attribute, and obtain the retrieval text corresponding to each target parameter.
  • the interactive content generating device further includes a product information pushing module.
  • the product information push module is used to push product recommendation information to the client in a preset product push format.
  • the interactive content generating device further includes a product information pushing module and a question text obtaining module.
  • the product information push module is used to determine whether the search text contains at least two target options.
  • the obtaining question text module is used to obtain the question text corresponding to the at least two target options if it contains at least two target options, call the information conversion model to convert the question text, obtain the current reply information, and push the current reply information to Client.
  • Each module in the above interactive content generating device can be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 9.
  • the computer device includes a processor, memory, network interface, and database connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer-readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer-readable instructions in the non-volatile storage medium.
  • the database of the computer equipment is used for data related to the interactive content generation method.
  • the network interface of the computer device is used to communicate with external terminals through a network connection.
  • the computer-readable instruction is executed by the processor to realize an interactive content generation method.
  • a computer device including a memory, a processor, and computer-readable instructions stored in the memory and capable of running on the processor.
  • the processor executes the computer-readable instructions to realize interactive content generation in the above-mentioned embodiments. Methods, such as S10 to S50 shown in FIG. 2.
  • the processor implements the functions of the modules/units of the interactive content generating apparatus in the foregoing embodiment when executing computer-readable instructions, for example, the functions of the modules 10 to 50 shown in FIG. 8. To avoid repetition, I won’t repeat them here.
  • one or more readable storage media storing computer readable instructions, the computer readable storage medium storing computer readable instructions, the computer readable instructions being executed by one or more processors
  • the one or more processors are executed to implement the interactive content generation method of the foregoing embodiment, for example, S10 to S50 shown in FIG. 2.
  • the computer-readable instruction is executed by the processor, the function of each module/unit in the interactive content generating apparatus in the above-mentioned apparatus embodiment is realized, for example, the function of the module 10 to the module 50 shown in FIG. 8. To avoid repetition, I won’t repeat them here.
  • the readable storage medium in this embodiment includes a nonvolatile readable storage medium and a volatile readable storage medium.
  • Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM random access memory
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDRSDRAM double data rate SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM synchronous chain (Synchlink) DRAM
  • RDRAM direct RAM
  • DRAM direct memory bus dynamic RAM
  • RDRAM memory bus dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Information Transfer Between Computers (AREA)

Abstract

一种交互内容生成方法、装置、计算机设备及存储介质，该方法包括：接收客户端发送的携带会话标识的当前轮次信息，基于当前轮次信息获取至少一个当前意图参数（S10）；基于会话标识查询会话记录数据库，获取与会话标识相对应的至少一个历史意图参数（S20）；采用预设的强化学习模型对至少一个当前意图参数和至少一个历史意图参数进行分析，获取目标意图，目标意图包括至少一个目标参数和参数优先级顺序（S30）；基于每一目标参数查询检索文本数据库，获取与每一目标参数相对应的检索文本（S40）；调用信息转化模型对每一检索文件进行转换，获取与每一检索文本相对应的当前回复信息，依据参数优先级顺序将至少一个当前回复信息推送给客户端（S50）。

Description

交互内容生成方法、装置、计算机设备及存储介质
本申请以2019年1月18日提交的申请号为201910047156.8，名称为“交互内容生成方法、装置、计算机设备及存储介质”的中国发明申请为基础，并要求其优先权。
技术领域
本申请涉及语义分析技术领域，尤其涉及一种交互内容生成方法、装置、计算机设备及存储介质。
背景技术
随着科学技术的不断发展，信息技术、计算机技术以及人工智能技术的引入，机器人的研究已经逐步走出工业领域，逐渐扩展到了医疗、保健、家庭、娱乐以及服务行业等领域。而人们对于机器人的要求也从简单重复的机械动作提升为具有拟人问答、自主性及与其他机器人进行交互的智能机器人，人机交互也就成为决定智能机器人发展的重要因素。
目前的聊天机器人大部分侧重于和用户进行单轮交互，使得机器人并不能很好地获取用户的真实意图，因为用户聊天的重要信息可能在用户之前的对话中。这种基于单轮的聊天机制，忽略了当前用户前几轮的聊天主题和场景分析，因此其返回的回复内容可能存在偏差甚至错误。如何得到更贴近用户真实意图的聊天内容成为亟待解决的问题。
发明内容
本申请实施例提供一种交互内容生成方法、装置、计算机设备及存储介质，以解决如何得到更贴近用户真实意图的聊天内容的问题。
一种交互内容生成方法，包括：
接收客户端发送的携带会话标识的当前轮次信息，基于当前轮次信息获取至少一个当前意图参数；
基于会话标识查询会话记录数据库，获取与会话标识相对应的至少一个历史意图参数；
采用预设的强化学习模型对至少一个当前意图参数和至少一个历史意图参数进行分析，获取目标意图，目标意图包括至少一个目标参数和参数优先级顺序；
基于每一目标参数查询检索文本数据库，获取与每一目标参数相对应的检索文本；
调用信息转化模型对每一检索文本进行转换，获取与每一检索文本相对应的当前回复信息，依据参数优先级顺序将至少一个当前回复信息推送给客户端。
一种交互内容生成装置，包括：
接收当前信息模块，用于接收客户端发送的携带会话标识的当前轮次信息，基于当前轮次信息获取至少一个当前意图参数；
获取历史参数模块，用于基于会话标识查询会话记录数据库，获取与会话标识相对应的至少一个历史意图参数；
获取目标意图模块，用于采用预设的强化学习模型对至少一个当前意图参数和至少一个历史意图参数进行分析，获取目标意图，目标意图包括至少一个目标参数和参数优先级顺序；
获取检索文本模块，用于基于每一目标参数查询检索文本数据库，获取与每一目标参数相对应的检索文本；
获取回复信息模块，用于调用信息转化模型对每一检索文本进行转换，获取与每一检索文本相对应的当前回复信息，依据参数优先级顺序将至少一个当前回复信息推送给客户端。
一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个 当前意图参数；
基于所述会话标识查询会话记录数据库，获取与所述会话标识相对应的至少一个历史意图参数；
采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，所述目标意图包括至少一个目标参数和参数优先级顺序；
基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本；
调用信息转化模型对每一所述检索文本进行转换，获取与每一所述检索文本相对应的当前回复信息，依据所述参数优先级顺序将至少一个所述当前回复信息推送给所述客户端。
一个或多个存储有计算机可读指令的可读存储介质，所述计算机可读存储介质存储有计算机可读指令，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如下步骤：
接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数；
基于所述会话标识查询会话记录数据库，获取与所述会话标识相对应的至少一个历史意图参数；
采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，所述目标意图包括至少一个目标参数和参数优先级顺序；
基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本；
调用信息转化模型对每一所述检索文本进行转换，获取与每一所述检索文本相对应的当前回复信息，依据所述参数优先级顺序将至少一个所述当前回复信息推送给所述客户端。
本申请的一个或多个实施例的细节在下面的附图及描述中提出。本申请的其他特征和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对本申请实施例的描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本申请一实施例中交互内容生成方法的应用环境示意图；
图2是本申请一实施例中交互内容生成方法的流程图；
图3是本申请一实施例中获取目标意图的实现过程示意图；
图4是本申请一实施例中交互内容生成方法的另一流程图；
图5是本申请一实施例中交互内容生成方法的另一流程图；
图6是本申请一实施例中交互内容生成方法的另一流程图；
图7是本申请一实施例中交互内容生成方法的另一流程图；
图8是本申请一实施例中交互内容生成装置的示意图；
图9是本申请一实施例中计算机设备的示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
本申请实施例提供的交互内容生成方法，可应用在如图1的应用环境中，该交互内容生 成方法应用在交互内容生成系统中，该交互内容生成系统包括客户端和服务器，其中，客户端通过网络与服务器进行通信。客户端又称为用户端，是指与服务器相对应，为客户提供本地服务的程序。该客户端可安装在但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备等计算机设备上。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现，用于接受用户通过客户端发送的当前轮次信息，生成聊天机器人对应的回复信息。
在一实施例中，如图2所示，提供一种交互内容生成方法，以该方法应用在图1中的服务器为例进行说明，包括如下步骤：
S10.接收客户端发送的携带会话标识的当前轮次信息，基于当前轮次信息获取至少一个当前意图参数。
其中，当前轮次信息是用户当前轮次输入客户端的用以表达用户意图的信息。进一步地，当前轮次信息可以包括以多种形式表达的意图信息，包括但不限于文字数据、语音信息或手势动作信息等，此处不作限定。
会话标识是用以区别服务器启动的不同会话场景的标识。因服务器可与若干客户端启动若干会话场景，为了区别每一会话场景以进行会话内容分析，服务器需要给每一会话场景设置对应的会话标识。
意图是用户在当前轮次信息中表达的目的，可分为“显式意图”和“隐式意图”，它们两者间的差异十分明显：
(一)显式意图：
所谓显式意图，也就是在用户的话语中明确出现了表达意图的一类词汇，比如：“希望”、“想要”、“需要”等等。对于聊天机器人来说这种显示意图的判断难度较低，只需要识别出这些固定的意图词汇，再与句子中的其他成分作联系即可。比如用户输入“我想预定到北京的机票”，聊天机器人可以识别出意图词汇“想”，从而得到该句的意图参数为我预定机票，机票行程包括从当前位置到北京。
(二)隐式意图：
隐式意图与显式意图相反，用户的话语没有出现直接反映意图的词汇，需要聊天机器人根据文字数据来判断用户意图。对于聊天机器人来说隐式意图比较难以判断，目前用得比较多的一种方法是先把隐式意图转化成显式意图再处理。比如用户在表达“我饿了”这种隐式意图时，先转化成对应的“我想吃东西”这种显式意图形式，然后再让聊天机器人按照显式意图进行处理。
当前意图参数是将当前轮次信息转换为对应的文字数据后，从该文字数据中提取的含有实际意义的动词和名词(去停用词)，以最简洁的方式表达的用户目的。其中，停用词主要包括英文字符、数字、数学字符、标点符号及使用频率高的功能词等。比如，文字数据为“今天天气真不错啊”，从该文字数据中提取的当前意图参数包括：今天、天气和不错(去除停用词“真”和“啊”)。
步骤S10中，服务器通过客户端发送的当前轮次信息可提取出至少一个当前意图参数，过滤掉当前轮次信息中无实际意义的功能词，可直接获取当前轮次信息中有效的当前意图参数，为后续结合历史意图参数获得用户的真实意图准备技术基础。
S20.基于会话标识查询会话记录数据库，获取与会话标识相对应的至少一个历史意图参数。
其中，会话记录数据库是服务器基于每一会话场景(也即每一会话标识)保存的会话记录集合。
历史意图参数是与当前意图参数对应的，基于同一会话标识的在当前轮次会话之前保存的所有轮次的意图参数。
具体地，服务器基于同一会话标识都在会话记录数据库中对应记录当前轮次对话产生的当前意图参数。可以理解地，每一当前意图参数被存储到会话记录数据库中即形成历史意图 参数，以备后续服务器可基于同一会话标识匹配出对应的所有历史意图参数进行真实意图分析。
步骤S20中，服务器可基于会话标识在会话记录数据库中直接获取对应的所有历史意图参数，无需服务器重新提取所有历史会话中的历史意图参数，加快服务器分析用户真实意图的处理速度。
S30.采用预设的强化学习模型对至少一个当前意图参数和至少一个历史意图参数进行分析，获取目标意图，目标意图包括至少一个目标参数和参数优先级顺序。
其中，强化学习模型是预设于服务器的，用以将输入该模型的所有历史意图参数和当前意图参数进行分析，获取分析结果作为目标意图的模型。于本实施例，服务器可采用LSTM(Long Short-Term Memory，长短期记忆网络)模型和softmax分类器的组合作为强化学习模型。
于本实施例针对聊天场景可设定多级目标意图，比如一级目标意图包括聊天和提问；一级目标意图又可继续划分多个二级目标意图，对于一级目标意图中的聊天意图可继续划分为生活、工作和休闲；基于二级目标意图，可继续划分为三级目标意图，根据场景需要继续细化等。
具体地，如图3所示，服务器通过文字数据获取目标意图的实现过程如下：
1.将文字数据进行预处理，包括去除语料的标点符号，去除停用词(无实际含义的词)等。
2.将预处理后的文字数据，采用word2vec工具生成词向量。
其中，word2vec工具是一个NLP(Natural Language Processing，自然语言处理)工具，它可以将所有的自然语言中的字词向量化，转为计算机可以理解的稠密向量(Dense Vector)，用以定量地度量词与词之间的关系，挖掘词与词之间的联系。可以理解地，对于相似的词，其对应的词向量也相近。
3.采用LSTM模型对词向量进行特征提取。
其中，LSTM模型可解决自然语言处理统计方法只能考虑最近n个词语而忽略更久前词语的问题，旨在发现词和词之间的相关关系，在数据分析中增加时间文本内容，记住之前发生了什么，然后应用于神经网络，观察与神经网络接下来所发生的事情之间的联系，从而得出目标意图。
LSTM的特点就是在RNN(Recurrent Neural Network，循环神经网络)模型以外添加了各层的阀门节点，如图4所示。阀门节点有3类：遗忘阀门(forget gate)，输入阀门(input gate)和输出阀门(output gate)。这些阀门节点可以打开或关闭，用于将判断RNN模型的记忆态(之前网络的状态)在上一输出层的结果是否达到阈值从而加入到当前该层的计算中。
阀门节点利用sigmoid函数将在RNN模型中该阀门节点对应的上一输出层运算结果作为输入计算；如果计算结果达到阈值则将该阀门节点输出与当前层的计算结果相乘作为下一层的输入；如果没有达到阈值则将该阀门节点对应的上一输出层运算结果遗忘掉。
LSTM模型的记忆功能就是由这些阀门节点实现的。当阀门节点打开的时候，前面RNN模型的训练结果就会关联到当前的RNN模型进行计算，而当阀门节点关闭的时候之前的计算结果就不再影响当前的计算。因此，通过调节阀门节点的开关可以实现历史意图对最终获取目标意图的影响。当不希望历史意图对之后的意图分析产生影响，比如自然语言处理中开始分析新段落或新章节，则关掉阀门节点即可。
4.采用softmax完成意图分类工作。
最后，将LSTM模型对词向量进行特征提取后，在神经网络中引入softmax分类器可定义一个新的输出层(目标意图概率分布)。基于神经网络的LSTM模型的输出层不一定是一个概率分布层，所以可给LSTM模型在最后引入softmax分类器，softmax分类器能够作为一个额外的处理层把神经网络的输出(LSTM模型的最后输出层)变成一个概率分布，也就是每一个输出都是0至1之间的小数，并且所有输出的结果之和为1。举例说明，若LSTM模型的最后输出层为y1,y2,y3和y4，服务器可采用softmax分类器对该LSTM模型的最后输出层进行处 理，采用如下公式：
Figure PCTCN2019120595-appb-000001
通过softmax分类器可以计算出y′ 1y′ 2y′ 3y′ 4的值，其中，y′ i是每一目标意图的输出概率。可以理解地，y′ i中数值最大的输出对应的目标意图就是当前轮次信息对应的目标意图。比如，步骤S10获得文字数据“我最近看了姜文导演的新电影邪不压正”，经步骤S20中预设的意图识别模型进行识别后获得的目标意图是：用户讨论电影，其中，电影参数包括：电影名为邪不压正。
进一步地，本实施例还可为目标参数设置参数优先级顺序，也即给动词后的指定动作的对象对应的名词设置较高的参数优先级顺序，给动作执行人设置次级参数优先级顺序。对于多数讨论场景，人们谈论事件的重点在于动作执行的对象，也即动作执行的对象的参数优先级顺序一般高于动作执行人。给不同的目标参数设置参数优先级顺序，利于后续基于不同的参数优先级顺序进行概念扩展。
或者，服务器还可设置将包括目标参数个数最多的目标意图的参数优先级的值也最高。比如，用户通过客户端输入的文字数据为“查询龙岗租房信息，面积20平米，价格便宜的”。服务器基于该文字数据可分析该文字数据的目标意图是龙岗租房，将该文字数据进行拆分分析：
1)“查询龙岗租房信息”包括的目标参数是龙岗租房；
2)“查询龙岗租房信息，面积20平米”包括的目标参数是龙岗租房和面积；
3)“查询龙岗租房信息，面积20平米，价格便宜的”包括的目标参数是龙岗租房、面积和价格。
由上述分析可以看出，第三组包括的目标参数最多，也即第三组的参数优先级顺序为3，依次类推，第二组的参数优先级顺序为2，第一组的参数优先级顺序为1。服务器基于参数优先级顺序进行回复时，可按参数优先级顺序的从大到小的顺序进行查找，也即将第三组查询到的租房结果首先返回给客户端。
举例说明在本实施例中，服务器通过目标意图获取目标参数的实现过程：
比如，文字数据“我最近看了姜文导演的新电影邪不压正”，经步骤S30的识别后可确认该文字数据对应的目标意图为动作人谈休闲活动。在动作人谈休闲活动的目标意图模板中可获取该模板的目标参数为“动作人”和“休闲活动”。在文字数据“我最近看了姜文导演的新电影邪不压正”中提取与上述目标参数分别对应的词语：动作人对应“我”(也即用户)，休闲活动“电影邪不压正”。其中，动作人对应的参数优先级顺序为1级，休闲活动对应的参数优先级顺序为2级，则在后续将目标参数作为回复信息回复给客户端时，选取参数优先级顺序的值最大的目标参数先回复给客户端。
步骤S30中，服务器能够通过预设的强化学习模型识别用户的目标意图，精准地判断出用户的需求，并基于目标参数对应的参数优先级顺序最大的优先回复，利于后续基于该目标参数得到相应的回复信息，准确地围绕用户的关注点推进聊天，以保持聊天机器人和用户之间的聊天黏性和实用性。
S40.基于每一目标参数查询检索文本数据库，获取与每一目标参数相对应的检索文本。
其中，联网数据库即为在线检索库，比如，百度、搜狗、谷歌或360搜索等。检索文本就是与目标参数对应的回复文本，比如，目标参数为“今天天气”，服务器通过联网数据库查询在线天气可获“今天天气”对应的天气查询结果，该天气查询结果就是检索文本。
步骤S40中，服务器可通过联网数据库获取目标参数对应的检索文本，给用户提供与当前轮次信息相关的文本信息，提高回复内容的准确性。
S50.调用信息转化模型对每一检索文本进行转换，获取与每一检索文本相对应的当前回 复信息，依据参数优先级顺序将至少一个当前回复信息推送给客户端。
具体地，本步骤中的信息转化模型是将文本信息转化为传递给用户的回复信息的形式的模型。比如，用户以语音形式输入当前轮次信息，则本步骤采用语音转化模型将检索文本转化为语音信息输出给用户，或者也可以按客户端的喜好设置直接以文本形式直接输出给客户端。
信息转化模型是将文本数据表达的含义与特定表达方式进行相互转化的模型，比如，将语音信息转化为对应的文字数据，将文字数据转化为对应的手语动作等。于本实施例，信息转化模型可基于多种表达方式进行转换，因此其包括多个转化模型。比如，将语音信息转化为文字信息的语音转化模型，比如RNN-HMM(Recurrent Neural Network-Hidden Markov Model即循环神经网络-隐马尔科夫)模型或LSTM-HMM(Long Short-Term Memory，长短期记忆网络-隐马尔科夫)模型等机器声学模型，或者将手势信息转化为文字信息的手势转化模型，比如FLDCRFs(Fuzzy based Latent dynamic Condition Random Fields,模糊隐动态条件随机场)等机器手势识别模型。应当理解地，本实施例中RNN-HMM机器声学模型、LSTM-HMM机器声学模型和FLDCRFs机器手势识别模型等机器学习模型为公知技术，此处不再赘述。
举例说明步骤S60中将检索文本转化为当前回复信息的实现过程，在步骤S10中，若用户通过客户端的麦克风输入当前轮次信息(语音信息)：“我最近看了姜文导演的新电影邪不压正”。服务器通过客户端获取麦克风采集的语音格式的当前轮次信息，因此可确定当前轮次信息为语音信息，即可采用RNN-HMM机器声学模型将该语音信息转化为“我最近看了姜文导演的新电影邪不压正”这一文字数据。步骤S60中，为了保持与用户交流方式的一致性，提高聊天趣味性，聊天计算机可将步骤S50得到的检索文本同样采用RNN-HMM机器声学模型，将检索文本转化为语音信息(当前回复信息)返回给客户端。
步骤S50中，服务器可按照参数优先级顺序将检索文本对应的当前回复信息返回给客户端，利于基于用户的关注点直接回复，及时给客户端返回准确地可供参考的当前回复信息，以提高聊天机器人和用户之间的聊天内容的关联性；服务器可根据场景需要设置为输出的回复信息的形式与用户输入的当前轮次信息的方式一致，保持聊天习惯一致性，提高用户与聊天机器人的聊天趣味性。
本实施例提供的交互内容生成方法中，服务器通过接收客户端发送的当前轮次信息，分析得到至少一个当前意图参数，将当前意图参数结合会话记录数据库获取至少一个历史意图参数后得到目标意图和对应的回复信息，可保障聊天机器人及时生成准确的回复信息回复给客户端，避免仅依靠当前意图参数获得不准确甚至无关的回复信息，提升聊天机器人和客户端之间的互动关联性和可靠性。
在一具体实施方式中，如图4所示，步骤S10中，即接收客户端发送的当前轮次信息，包括：
S11.接收客户端发送的当前轮次信息，调用信息转化模型对当前轮次信息进行识别，获取文字数据。
其中，文字数据就是以特定方式发送的当前轮次信息对应含义的文字信息，比如，用户通过客户端的摄像头发送点头动作作为当前轮次信息，服务器可通过调用信息转化模型(于本实施例为动作识别模型)将点头动作对应转化为文字数据“同意”。
步骤S11中，服务器通过将接受到的客户端发送的当前轮次信息转换为文字数据，利于服务器后续基于对文字数据进行进一步处理，以获取与当前轮次信息对应的回复信息返回给客户端。
S12.采用预设语言处理模型对文字数据进行解析处理，获取至少一个当前意图参数。
其中，本实施例的预设语言处理模型可采用word2vec，一个NLP(Natural Language Processing，自然语言处理)工具，它可以将所有的自然语言中的字词向量化，转为计算机可以理解的稠密向量(Dense Vector)，用以定量地度量词与词之间的关系，挖掘词与词之间的联系。可以理解地，对于相似的词，其对应的词向量也相近。
本步骤中采用预设语言处理模型对文字数据进行解析处理，获取至少一个当前意图参数的实现过程在步骤S10中已进行详细描述，为了避免重复，此处不再赘述。
步骤S12中，服务器可采用预设语言处理模型解析文字数据获得至少一个当前意图参数，可提取出用户在当前轮次表达的最简洁的当前意图，利于后续服务器基于该当前意图结合所有历史意图参数获取用户的真实意图。
步骤S11至S12中，服务器通过将接受到的客户端发送的当前轮次信息转换为文字数据，利于服务器后续基于对文字数据进行进一步处理，以获取与当前轮次信息对应的回复信息返回给客户端。服务器可采用预设语言处理模型解析文字数据获得至少一个当前意图参数，可提取出用户在当前轮次表达的最简洁的当前意图，利于后续服务器基于该当前意图结合所有历史意图参数获取用户的真实意图。
在一具体实施方式中，如图5所示，步骤S30中，即采用预设的强化学习模型对至少一个当前意图参数和至少一个历史意图参数进行分析，获取目标意图，包括：
S31.基于至少一个当前意图参数和至少一个历史意图参数，分别获取当前词向量矩阵和历史词向量矩阵。
其中，词向量其实是将词映射到一个语义空间，得到的语义映射矩阵。举例说明，中心词A映射出周边词BCDAEFG，得到一个参数矩阵W1，然后将中心词L映射到周边词BCDLEFG，得到参数矩阵W2。若W2和W1距离相近，则说明A和L能映射到同样的周边词，可能是近义词。
词向量训练的预处理步骤：
1.对输入的文本(当前意图参数或历史意图参数)生成一个词汇表，统计词汇表中每个词的词频，按照词频从高到低排序，取最频繁的V个词，构成一个词汇表。每个词存在一个维度是V的one-hot向量。其中，one-hot向量用以表征向量中每一个元素都关联着词库中的一个单词，指定词的向量表示为：其在向量中对应的元素设置为1，其他的元素设置为0(如果该词在词汇表中出现过，则向量中词汇表中对应的位置为1，其他位置全为0)。如果词汇表中不出现，则向量为全0。
2.将输入文本的每个词都生成一个one-hot向量，保留每个词的原始位置，因为是上下文相关的。
3.确定词向量的维数N。
具体地，本实施例可采用CBOW模型(连续词袋模型，将词语的上下文作为输入，来预测这个词语本身)获取词向量，实现步骤如下：
1.确定窗口大小window，对每个词生成2*window个训练样本，(i-window,i)，(i-window+1,i)，...，(i+window-1,i)，(i+window,i)。
2.确定batch_size，batch_size的大小须是2*window的整数倍，以确保每个batch包含了一个词汇对应的所有样本。
3.训练算法有两种：层次Softmax和Negative Sampling。
4.神经网络迭代训练一定次数，得到输入层到维数为N的隐藏层的参数矩阵，矩阵中每一行的转置即是对应词的维数为V的词向量矩阵。
步骤S31中，服务器可分别获取当前意图参数和历史意图参数对应的当前词向量矩阵和历史词向量矩阵，为后续服务器基于词向量矩阵获取目标意图准备技术基础。
S32.采用预设的强化学习模型对当前词向量矩阵和历史词向量矩阵进行分析，获取目标意图。
具体地，本步骤中可采用步骤S30中的LSTM(Long Short-Term Memory，长短期记忆网络)模型和softmax分类器的组合作为强化学习模型，对当前词向量矩阵和历史词向量矩阵进行分析进行解析处理，获取目标意图。
将当前词向量矩阵和历史词向量矩阵输入到预设的强化学习模型的输入阀门，具体的实现过程已在前述步骤S30中进行详细描述，为了避免重复，此处不再赘述。
步骤S32中，服务器能够通过预设的强化学习模型识别用户的目标意图，精准地判断出用户的需求，获取与用户需求更贴合的聊天内容，减少聊天机器人和用户之间答非所问的次数。
步骤S31至S32中，服务器可分别获取当前意图参数和历史意图参数对应的当前词向量矩阵和历史词向量矩阵，为后续服务器基于词向量矩阵获取目标意图准备技术基础。服务器能够通过预设的强化学习模型识别用户的目标意图，精准地判断出用户的需求，获取与用户需求更贴合的聊天内容，减少聊天机器人和用户之间答非所问的次数。
在一具体实施方式中，如图6所示，步骤40中，基于每一目标参数查询联网数据库，获取与每一目标参数相对应的检索文本，具体包括如下步骤：
S41.若目标参数的属性为产品属性，则基于每一目标参数查询与产品属性相对应的产品数据库，获取与每一目标参数相对应的产品推荐信息作为检索文本。
其中，产品属性是内容提供商通过服务器给每一目标参数标记的事物属性。其中，内容提供商是提供聊天机器人所在服务器的提供商，该提供商可同时售卖其它实物产品或提供各种服务。
可以理解地，当本聊天机器人应用于内容提供商的营销场景，聊天机器人可在用户沟通的恰当时机对内容提供商提供的营销产品或服务进行推介，利于采用该聊天机器人系统的内容提供商提高产品或服务销售转化率。比如，目标参数为“保险”，若保险属于内容提供商的自由产品，内容提供商可给“保险”标注为产品属性。
具体地，服务器可将每一目标意图模板中的实物名称或服务名称与内容提供商自身提供的实物或服务类型进行关联，用以在服务器获取到的目标参数包括其中的实物名称或服务名称时，内容提供商可根据关联关系给用户推出对应的产品或服务介绍。
产品数据库是保存内容提供商可提供的所有产品或服务的数据库。产品推荐信息是产品数据库中给每一产品或服务记录的产品内容介绍。
优选的，产品推荐信息可包括产品概要介绍和产品详细介绍，用以服务器首先推送给客户端产品概要介绍。当服务器接收到客户端发送的产品详细介绍请求时，表明用户有意愿在继续了解产品的详细情况时，可再继续给客户端发送产品对应的产品详细介绍。
步骤S41中，服务器可预先给每一目标参数标注是否属于产品属性的标签，利于服务器获取到目标参数即刻可进行判断，无需再对每一目标参数进行分析是否为产品属性，利于缩短服务器的判定时间。服务器可基于产品数据库获取目标参数对应的产品推荐信息，用以后续推送给用户，提升采用本聊天机器人的实际适用性和聊天内容可扩展性。
优选地，在步骤S41之后，即在获取与每一目标参数相对应的产品推荐信息作为检索文本之后，交互内容生成方法还包括:将产品推荐信息按预设产品推送格式推送给客户端。
其中，预设产品推送格式是服务器预先设置的将产品在聊天界面推送给客户端的格式，比如，服务器可设定推送产品名列表，或者每一产品对应产品简单介绍等。
进一步地，当服务器接收到客户端选中的产品名，可将该产品名与步骤S51中建立的产品详细介绍关联，以将产品详细介绍再推送给客户端，避免初次即推送给客户端的用户不感兴趣的产品的详细介绍。
本步骤中，服务器可按预设产品推送格式将产品推荐信息发送给客户端，保持聊天内容涉及到产品或服务时的格式一致性。
S42.若目标参数的属性不为产品属性，基于每一目标参数查询联网数据库，获取与每一目标参数相对应的检索文本。
其中，联网数据库即为在线检索库，比如，百度、搜狗、谷歌或360搜索等。检索文本就是与目标参数对应的回复文本，比如，目标参数为“今天天气”，服务器通过联网数据库查询在线天气可获“今天天气”对应的天气查询结果，该天气查询结果就是检索文本。
步骤S42中，服务器可通过联网数据库获取目标参数对应的检索文本，给用户提供与当前轮次信息相关的检索文本，提高回复内容的准确性；同时，服务器无需在本地保存与目标参数对应的检索文本，可节省服务器本地的存储空间。
步骤S41至S42中，服务器可预先给每一目标参数标注是否属于产品属性的标签，利于服务器获取到目标参数即可进行判断，无需再对每一目标参数进行分析是否为产品属性，利于缩短服务器的判定时间。服务器可基于产品数据库获取目标参数对应的产品推荐信息，用以后续推送给用户，提升采用本聊天机器人的实际适用性和聊天内容可扩展性。服务器可通过联网数据库获取目标参数对应的检索文本，给用户提供与当前轮次信息相关的检索文本，提高回复内容的准确性；同时，服务器无需在本地保存与目标参数对应的检索文本，可节省服务器本地的存储空间。
在一具体实施方式中，如图7所示，在步骤S50之前，即在所述调用信息转化模型对每一所述检索文本进行转换之前，所述交互内容生成还包括：
S501.获取用户个人信息，基于用户个人信息生成用户画像。
其中，用户个人信息是服务器收集的静态数据和动态数据，其中，静态数据是用户在服务器注册时主动输入的长时间内不会变化的个人信息，比如，用户的地域、年龄、性别、文化、职业和收入等；动态数据是服务器与用户互动过程中对用户行为进行分析得到的数据，比如，生活习惯或消费习惯等。总体来讲，为了获取用户画像需要用到如下用户个人信息：
(1)人口属性：包括性别、年龄等基本信息；
(2)兴趣特征：浏览内容、收藏内容、阅读咨询、购买物品偏好等；
(3)消费特征：与消费相关的特征；
(4)位置特征：用户所处城市、所处居住区域、用户移动轨迹等；
(5)设备属性：使用的终端特征等；
(6)行为数据：访问时间、浏览路径等用户在网站的行为日志数据；
(7)社交数据：用户社交相关数据。
用户画像是服务器根据用户个人信息而抽象出来的高度精炼的标签，也即用户特征标识。通过给用户“贴”标签可以利用高度概括、容易理解的特征来描述用户，利于服务器根据标签(标准化信息)进一步处理。
具体地，基于用户个人信息生成用户画像的实现过程如下：
一、对用户个人信息中的某项特征进行抽象分类和概括形成标签，该标签的标签值具备可分类性。
例：将用户个人信息中“男”、“女”这类特征进行抽象概括，统称为“性别”，“性别”即一个标签；
二.穷举标签值(Tag Value)，以使该标签包括对应的所有可能的情况的值。
例：对于标签“性别”，其标签值可分为“男”、“女”和“未知”；
对于标签“年龄”，其标签值可分为“0-18”、“18-35”、“35-60”、“60-100”等。
三.构建用户画像(User Profile)。根据步骤一和步骤二创造的标签，提取用户个人信息中与每一标签对应的标签值。
例：用户画像包括的标签有性别、年龄、手机品牌、居住地和爱好等。小明是用户画像的一个实例，小明的用户画像的输出结果为：“男“、“18-35”、“iPhone”、“北京”和“足球”。
步骤S501中，服务器可基于用户个人信息构建用户画像，给后续服务器基于该用户画像分析用户的用户性格准备技术基础。
S502.分析用户画像，获取与用户画像相对应的用户性格。
其中，用户性格是一个人对现实的稳定的态度，以及与这种态度相应的，习惯化了的行为方式中表现出来的人格特征。性格一经形成便比较稳定，但是并非一成不变，而是可塑性的。性格不同于气质，更多体现了人格的社会属性，个体之间的人格差异的核心是性格的差 异。
具体地，通过对用户画像进行分析和挖掘，还可以揭示用户的内心需求和用户性格等潜在属性。比如，用户画像中有关购物标签包括平均浏览时间和平均对比同类次数，三个用户的网上购物标签记录情况如下：
  平均浏览时间(分钟) 平均对比次数
用户A 7 3
用户B 15 12
用户C 25 20
购物性格参考模板(基于实际情况统计生成)：
冲动型：平均浏览时间少于10分钟，平均对比次数少于5次。
理性型：平均浏览时间介于10分钟至20分钟，平均对比次数介于5次至20次。
犹豫型：平均浏览时间大于20分钟，平均对比次数大于20次。
服务器基于三个用户的购物标签和购物性格参考模板进行对比可得如下结果：
用户A总在短时间(平均浏览时间少于10分钟)内比较少量的商品(平均对比次数少于5次)下单，则用户A的购物性格为冲动型。
用户B总在适量时间内(平均浏览时间介于10分钟至20分钟)比较少量同类商品(平均对比次数介于5次至20次)后下单，则用户B的购物性格为理性型。
用户C总在长时间(平均浏览时间大于20分钟)大量地浏览很多商品(平均对比次数大于20次)后下单，则用户C的购物性格为犹豫型。
同样地，服务器还可对用户画像进行分析获得用户的聊天性格，聊天性格包括：兴奋、平静或低沉等。需要补充的是，服务器还可通过分析用户输入的当前轮次信息中的声音特质结合聊天性格参考模板来判断用户性格。其中，声音特征的聊天性格参考模板(基于实际情况统计生成)如下：
1)掺杂呼吸且脆弱的声音：如果用户是男性，属于年轻艺术家的类型；如果用户是女性，虽性格偏女性化，且集美丽、小巧、开朗于一身，这类女性比较容易兴奋。
2)无力的声音：如果用户是男性，无特定性格；如果用户是女性，则具有社交能力强、感性、幽默等特性。
3)平淡的声音：用户无论男女，性格特征均为偏男性化、精神状态欠佳、冷淡、畏缩不前，当前状态不太积极。
4)声调低且粗犷的声音：如果用户是男性，应该拥有敏锐的观察力，且现实、圆滑、成熟、干练、适应性强；如果用户是女性，性格特质主要表现为喜欢偷懒或病怏怏等。
5)爽朗的声音：如果用户是男性，性格干练且自尊心强；如果用户是女性，则性格活泼、善于社交、自尊心强，缺乏幽默感。
6)语速快：用户无论男女，该类人群的性格均较活泼，且社交能力强。
7)抑扬顿挫且洪亮的声音：用户无论男女，该类人群精力旺盛。
步骤S502中，服务器可对步骤S601中获取的用户画像根据实际经验进行分析，获得用户对应的用户性格，为后续服务器基于不同的用户性格采用不同的回复模式准备技术基础。
S503.基于用户性格查询预置回复模式库，获取与用户性格对应的目标回复模式。
其中，回复模式库是预置于服务器的，针对不同的用户性格的回复信息的模式组成的数据库。
目标回复模式是针对用户性格进行回复的聊天模式，本实施例主要应用于语音聊天，因 此回复模式即为聊天机器人与用户聊天时的语气回复模式。
具体地，根据心理学研究聊天时语气同步的双方，更加容易被对方接受。也就是说如果用户说话语气慢条斯理，若聊天机器人跟着用户的节奏聊天，可增加用户与聊天机器人继续聊天的意愿。基于此，服务器可根据用户性格搭配对应的聊天语气，以进一步增强用户和聊天机器人之间的聊天趣味性和情感交互性。
进一步地，服务器可将对每一用户性格配置一种语气回复模式，比如，对于热情、兴奋或积极等用户性格的语气回复模式为快速且热情的语气回复模式；对于冷静、平淡或成熟的用户性格的语气回复模式为中速且平静的语气回复模式等。
步骤S503中，服务器可根据用户性格在预置回复模式库中匹配对应的目标回复模式，以进一步增强用户和聊天机器人之间的聊天趣味性和情感交互性。
S504.将目标回复模式添加到信息转化模型，以更新信息转化模型。
具体地，服务器将步骤S603获得的目标回复模式(可以为语气回复模式)，比如快速且热情的语气回复模式等发送给信息转化模型，以使信息转化模型将概念文本转换为与目标回复模式相匹配的当前回复信息。进一步地，信息转化模型可通过调整语速和说话的音调实现获取与目标回复模式相匹配的当前回复信息。
步骤S504中，服务器可结合目标回复模式和信息转化模型获得更新的信息转化模型，用以调整与目标回复模式相匹配的语气，将对应的回复信息发送给客户端，增强聊天机器人和用户之间的情感交互性，提高聊天内容的有效性。
步骤S501至S504中，服务器可根据实际经验对用户画像进行分析，获得用户对应的用户性格，并根据用户性格在预置回复模式库中匹配对应的目标回复模式，以进一步增强用户和聊天机器人之间的聊天趣味性和情感交互性。服务器可将调整后与目标回复模式相匹配的当前回复信息发送给客户端，增强聊天机器人和用户之间的情感交互性，提高聊天内容的有效性。
本实施例提供的交互内容生成方法中，服务器通过接收客户端发送的当前轮次信息，分析得到至少一个当前意图参数，将当前意图参数结合会话记录数据库获取至少一个历史意图参数后得到目标意图和对应的回复信息，可保障聊天机器人及时生成准确的回复信息回复给客户端，避免仅依靠当前意图参数获得不准确甚至无关的回复信息，提升聊天机器人和客户端之间的互动关联性和可靠性。
进一步地，服务器通过将接受到的客户端发送的当前轮次信息转换为文字数据，利于服务器后续基于对文字数据进行进一步处理，以获取与当前轮次信息对应的回复信息返回给客户端。服务器可采用预设语言处理模型解析文字数据获得至少一个当前意图参数，可提取出用户在当前轮次表达的最简洁的当前意图，利于后续服务器基于该当前意图结合所有历史意图参数获取用户的真实意图。
进一步地，服务器可分别获取当前意图参数和历史意图参数对应的当前词向量矩阵和历史词向量矩阵，为后续服务器基于词向量矩阵获取目标意图准备技术基础。服务器能够通过预设的强化学习模型识别用户的目标意图，精准地判断出用户的需求，获取与用户需求更贴合的聊天内容，减少聊天机器人和用户之间答非所问的次数。
进一步地，服务器可预先给每一目标参数标注是否属于产品属性的标签，利于服务器获取到目标参数即刻可进行判断，无需再对每一目标参数进行分析是否为产品属性，利于缩短服务器的判定时间。服务器可基于产品数据库获取目标参数对应的产品推荐信息，用以后续推送给用户，提升采用本聊天机器人的实际适用性和聊天内容可扩展性。服务器可通过联网数据库获取目标参数对应的检索文本，给用户提供与当前轮次信息相关的检索文本，提高回复内容的准确性；同时，服务器无需在本地保存与目标参数对应的检索文本，可节省服务器本地的存储空间。
进一步地，服务器可根据实际经验对用户画像进行分析，获得用户对应的用户性格，并根据用户性格在预置回复模式库中匹配对应的目标回复模式，以进一步增强用户和聊天机器 人之间的聊天趣味性和情感交互性。服务器可将调整后与目标回复模式相匹配的当前回复信息发送给客户端，增强聊天机器人和用户之间的情感交互性，提高聊天内容的有效性。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
在一实施例中，提供一种交互内容生成装置，该交互内容生成装置与上述实施例中交互内容生成方法一一对应。如图8所示，该交互内容生成装置包括接收当前信息模块10、获取历史参数模块20、获取目标意图模块30、获取目标意图模块30、获取检索文本模块40和获取回复信息模块50。各功能模块详细说明如下：
接收当前信息模块10，用于接收客户端发送的携带会话标识的当前轮次信息，基于当前轮次信息获取至少一个当前意图参数。
获取历史参数模块20，用于基于会话标识查询会话记录数据库，获取与会话标识相对应的至少一个历史意图参数。
获取目标意图模块30，用于采用预设的强化学习模型对至少一个当前意图参数和至少一个历史意图参数进行分析，获取目标意图，目标意图包括至少一个目标参数和参数优先级顺序。
获取检索文本模块40，用于基于每一目标参数查询检索文本数据库，获取与每一目标参数相对应的检索文本。
获取回复信息模块50，用于调用信息转化模型对每一检索文本进行转换，获取与每一检索文本相对应的当前回复信息，依据参数优先级顺序将至少一个当前回复信息推送给客户端。
优选地，接收当前信息模块10包括接收轮次信息单元11和获取意图参数单元12。
接收轮次信息单元11，用于接收客户端发送的当前轮次信息，调用信息转化模型对当前轮次信息进行识别，获取文字数据。
获取意图参数单元12，用于采用预设语言处理模型对文字数据进行解析处理，获取至少一个当前意图参数。
优选地，获取目标意图模块包括分别获取当前词向量矩阵和历史词向量矩阵单元、获取词性向量矩阵和获取目标意图单元。
分别获取当前词向量矩阵和历史词向量矩阵单元，用于基于至少一个当前意图参数和至少一个历史意图参数，分别获取当前词向量矩阵和历史词向量矩阵。
获取词性向量矩阵，用于基于当前意图参数和历史意图参数的词性标注，获取与当前词向量矩阵和历史词向量矩阵分别对应的当前词性向量矩阵和历史词性向量矩阵。
获取目标意图单元，用于采用预设的强化学习模型对当前词向量矩阵和历史词向量矩阵进行分析，获取目标意图。
优选地，获取检索文本模块包括获取产品信息单元和获取检索文本单元。
获取产品信息单元，用于若目标参数的属性为产品属性，则基于每一目标参数查询与产品属性相对应的产品数据库，获取与每一目标参数相对应的产品推荐信息作为检索文本。
获取检索文本单元，用于若目标参数的属性不为产品属性，基于每一目标参数查询联网数据库，获取与每一目标参数相对应的检索文本。
优选地，该交互内容生成装置还包括推送产品信息模块。
推送产品信息模块，用于将产品推荐信息按预设产品推送格式推送给客户端。
优选地，该交互内容生成装置还包括推送产品信息模块和获取问题文本模块。
推送产品信息模块，用于判断检索文本是否包含至少两个目标选项。
获取问题文本模块，用于若包含至少两个目标选项，则获取与至少两个目标选项相对应的问题文本，调用信息转化模型对问题文本进行转换，获取当前回复信息，将当前回复信息推送给客户端。
关于交互内容生成装置的具体限定可以参见上文中对于交互内容生成方法的限定，在此不再赘述。上述交互内容生成装置中的各个模块可全部或部分通过软件、硬件及其组合来实 现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图9所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于交互内容生成方法相关的数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种交互内容生成方法。
在一实施例中，提供一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机可读指令，处理器执行计算机可读指令时实现上述实施例交互内容生成方法，例如图2所示S10至步骤S50。或者，处理器执行计算机可读指令时实现上述实施例中交互内容生成装置的各模块/单元的功能，例如图8所示模块10至模块50的功能。为避免重复，此处不再赘述。
在一实施例中，一个或多个存储有计算机可读指令的可读存储介质，所述计算机可读存储介质存储有计算机可读指令，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行时实现上述实施例交互内容生成方法，例如图2所示S10至步骤S50。或者，该计算机可读指令被处理器执行时实现上述装置实施例中交互内容生成装置中各模块/单元的功能，例如图8所示模块10至模块50的功能。为避免重复，此处不再赘述。本实施例中的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，该计算机可读指令可存储于一非易失性可读存储介质也可以存储在易失性可读存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。
以上实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种交互内容生成方法，其特征在于，包括：
    接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数；
    基于所述会话标识查询会话记录数据库，获取与所述会话标识相对应的至少一个历史意图参数；
    采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，所述目标意图包括至少一个目标参数和参数优先级顺序；
    基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本；
    调用信息转化模型对每一所述检索文本进行转换，获取与每一所述检索文本相对应的当前回复信息，依据所述参数优先级顺序将至少一个所述当前回复信息推送给所述客户端。
  2. 如权利要求1所述的交互内容生成方法，其特征在于，所述接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数，包括：
    接收客户端发送的当前轮次信息，调用信息转化模型对所述当前轮次信息进行识别，获取文字数据；
    采用预设语言处理模型对所述文字数据进行解析处理，获取至少一个当前意图参数。
  3. 如权利要求1所述的交互内容生成方法，其特征在于，所述采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，包括：
    基于至少一个所述当前意图参数和至少一个所述历史意图参数，分别获取当前词向量矩阵和历史词向量矩阵；
    采用预设的强化学习模型对所述当前词向量矩阵和所述历史词向量矩阵进行分析，获取目标意图。
  4. 如权利要求1所述的交互内容生成方法，其特征在于，所述基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本，包括：
    若所述目标参数的属性为产品属性，则基于每一所述目标参数查询与所述产品属性相对应的产品数据库，获取与每一所述目标参数相对应的产品推荐信息作为检索文本；
    若所述目标参数的属性不为产品属性，基于每一所述目标参数查询联网数据库，获取与每一所述目标参数相对应的检索文本。
  5. 如权利要求4所述的交互内容生成方法，其特征在于，若所述目标参数的属性为产品属性，在所述获取与每一所述目标参数相对应的检索文本之后，所述交互内容生成方法还包括:
    将所述产品推荐信息按预设产品推送格式推送给所述客户端。
  6. 如权利要求1所述的交互内容生成方法，其特征在于，在所述调用信息转化模型对每一所述检索文本进行转换之前，所述交互内容生成方法还包括：
    获取用户个人信息，基于所述用户个人信息生成用户画像；
    分析所述用户画像，获取与所述用户画像相对应的用户性格；
    基于所述用户性格查询预置回复模式库，获取与所述用户性格对应的目标回复模式；
    将所述目标回复模式添加到所述信息转化模型，以更新所述信息转化模型。
  7. 一种交互内容生成装置，其特征在于，所述交互内容生成装置包括：
    接收当前信息模块，用于接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数；
    获取历史参数模块，用于基于所述会话标识查询会话记录数据库，获取与所述会话标识相对应的至少一个历史意图参数；
    获取目标意图模块，用于采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，所述目标意图包括至少一个目标参数和参数优先级顺序；
    获取检索文本模块，用于基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本；
    获取回复信息模块，用于调用信息转化模型对每一所述检索文本进行转换，获取与每一所述检索文本相对应的当前回复信息，依据所述参数优先级顺序将至少一个所述当前回复信息推送给所述客户端。
  8. 如权利要求7所述的交互内容生成装置，其特征在于，接收当前信息模块包括：
    接收轮次信息单元，用于接收客户端发送的当前轮次信息，调用信息转化模型对所述当前轮次信息进行识别，获取文字数据；
    获取意图参数单元，用于采用预设语言处理模型对所述文字数据进行解析处理，获取至少一个当前意图参数。
  9. 一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现如下步骤：
    接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数；
    基于所述会话标识查询会话记录数据库，获取与所述会话标识相对应的至少一个历史意图参数；
    采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，所述目标意图包括至少一个目标参数和参数优先级顺序；
    基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本；
    调用信息转化模型对每一所述检索文本进行转换，获取与每一所述检索文本相对应的当前回复信息，依据所述参数优先级顺序将至少一个所述当前回复信息推送给所述客户端。
  10. 如权利要求9所述的计算机设备，其特征在于，所述接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数，包括：
    接收客户端发送的当前轮次信息，调用信息转化模型对所述当前轮次信息进行识别，获取文字数据；
    采用预设语言处理模型对所述文字数据进行解析处理，获取至少一个当前意图参数。
  11. 如权利要求9所述的计算机设备，其特征在于，所述采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，包括：
    基于至少一个所述当前意图参数和至少一个所述历史意图参数，分别获取当前词向量矩阵和历史词向量矩阵；
    采用预设的强化学习模型对所述当前词向量矩阵和所述历史词向量矩阵进行分析，获取目标意图。
  12. 如权利要求9所述的计算机设备，其特征在于，所述基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本，包括：
    若所述目标参数的属性为产品属性，则基于每一所述目标参数查询与所述产品属性相对应的产品数据库，获取与每一所述目标参数相对应的产品推荐信息作为检索文本；
    若所述目标参数的属性不为产品属性，基于每一所述目标参数查询联网数据库，获取 与每一所述目标参数相对应的检索文本。
  13. 如权利要求12所述的计算机设备，其特征在于，若所述目标参数的属性为产品属性，在所述获取与每一所述目标参数相对应的检索文本之后，所述处理器执行所述计算机程序时还实现如下步骤:
    将所述产品推荐信息按预设产品推送格式推送给所述客户端。
  14. 如权利要求9所述的计算机设备，其特征在于，在所述调用信息转化模型对每一所述检索文本进行转换之前，所述处理器执行所述计算机程序时还实现如下步骤：
    获取用户个人信息，基于所述用户个人信息生成用户画像；
    分析所述用户画像，获取与所述用户画像相对应的用户性格；
    基于所述用户性格查询预置回复模式库，获取与所述用户性格对应的目标回复模式；
    将所述目标回复模式添加到所述信息转化模型，以更新所述信息转化模型。
  15. 一个或多个存储有计算机可读指令的可读存储介质，所述计算机可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如下步骤：
    接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数；
    基于所述会话标识查询会话记录数据库，获取与所述会话标识相对应的至少一个历史意图参数；
    采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，所述目标意图包括至少一个目标参数和参数优先级顺序；
    基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本；
    调用信息转化模型对每一所述检索文本进行转换，获取与每一所述检索文本相对应的当前回复信息，依据所述参数优先级顺序将至少一个所述当前回复信息推送给所述客户端。
  16. 如权利要求15所述的可读存储介质，其特征在于，所述接收客户端发送的携带会话标识的当前轮次信息，基于所述当前轮次信息获取至少一个当前意图参数，包括：
    接收客户端发送的当前轮次信息，调用信息转化模型对所述当前轮次信息进行识别，获取文字数据；
    采用预设语言处理模型对所述文字数据进行解析处理，获取至少一个当前意图参数。
  17. 如权利要求15所述的可读存储介质，其特征在于，所述采用预设的强化学习模型对至少一个所述当前意图参数和至少一个所述历史意图参数进行分析，获取目标意图，包括：
    基于至少一个所述当前意图参数和至少一个所述历史意图参数，分别获取当前词向量矩阵和历史词向量矩阵；
    采用预设的强化学习模型对所述当前词向量矩阵和所述历史词向量矩阵进行分析，获取目标意图。
  18. 如权利要求15所述的可读存储介质，其特征在于，所述基于每一所述目标参数查询检索文本数据库，获取与每一所述目标参数相对应的检索文本，包括：
    若所述目标参数的属性为产品属性，则基于每一所述目标参数查询与所述产品属性相对应的产品数据库，获取与每一所述目标参数相对应的产品推荐信息作为检索文本；
    若所述目标参数的属性不为产品属性，基于每一所述目标参数查询联网数据库，获取与每一所述目标参数相对应的检索文本。
  19. 如权利要求18所述的可读存储介质，其特征在于，若所述目标参数的属性为产品属性，在所述获取与每一所述目标参数相对应的检索文本之后，所述计算机可读指令被一 个或多个处理器执行时，使得所述一个或多个处理器还执行如下步骤:
    将所述产品推荐信息按预设产品推送格式推送给所述客户端。
  20. 如权利要求15所述的可读存储介质，其特征在于，在所述调用信息转化模型对每一所述检索文本进行转换之前，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器还执行如下步骤：
    获取用户个人信息，基于所述用户个人信息生成用户画像；
    分析所述用户画像，获取与所述用户画像相对应的用户性格；
    基于所述用户性格查询预置回复模式库，获取与所述用户性格对应的目标回复模式；
    将所述目标回复模式添加到所述信息转化模型，以更新所述信息转化模型。
PCT/CN2019/120595 2019-01-18 2019-11-25 交互内容生成方法、装置、计算机设备及存储介质 Ceased WO2020147428A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910047156.8 2019-01-18
CN201910047156.8A CN109857848A (zh) 2019-01-18 2019-01-18 交互内容生成方法、装置、计算机设备及存储介质

Publications (1)

Publication Number Publication Date
WO2020147428A1 true WO2020147428A1 (zh) 2020-07-23

Family

ID=66895125

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/120595 Ceased WO2020147428A1 (zh) 2019-01-18 2019-11-25 交互内容生成方法、装置、计算机设备及存储介质

Country Status (2)

Country Link
CN (1) CN109857848A (zh)
WO (1) WO2020147428A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115525740A (zh) * 2021-06-24 2022-12-27 Tcl科技集团股份有限公司 对话应答语句的生成方法、装置、电子设备及存储介质
CN117421398A (zh) * 2023-10-09 2024-01-19 北京百度网讯科技有限公司 人机交互方法、装置、设备以及存储介质

Families Citing this family (42)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109857848A (zh) * 2019-01-18 2019-06-07 深圳壹账通智能科技有限公司 交互内容生成方法、装置、计算机设备及存储介质
CN110222185A (zh) * 2019-06-13 2019-09-10 哈尔滨工业大学（深圳） 一种关联实体的情感信息表示方法
CN110297544B (zh) * 2019-06-28 2021-08-17 联想(北京)有限公司 输入信息响应方法及装置、计算机系统和可读存储介质
CN110489519B (zh) * 2019-07-05 2023-07-07 深圳追一科技有限公司 基于会话预测模型的会话方法及相关产品
CN110555165B (zh) * 2019-07-23 2023-04-07 平安科技（深圳）有限公司 信息识别方法、装置、计算机设备和存储介质
CN110390108B (zh) * 2019-07-29 2023-11-21 中国工商银行股份有限公司 基于深度强化学习的任务型交互方法和系统
CN110413788B (zh) * 2019-07-30 2023-01-31 携程计算机技术（上海）有限公司 会话文本的场景类别的预测方法、系统、设备和存储介质
CN112445906A (zh) * 2019-08-28 2021-03-05 北京搜狗科技发展有限公司 一种生成回复消息的方法及装置
CN110704641B (zh) * 2019-10-11 2023-04-07 零犀(北京)科技有限公司 一种万级意图分类方法、装置、存储介质及电子设备
CN110909137A (zh) * 2019-10-12 2020-03-24 平安科技（深圳）有限公司 基于人机交互的信息推送方法、装置和计算机设备
US11651250B2 (en) 2019-11-20 2023-05-16 International Business Machines Corporation Automatically generated conversation output
CN112825125A (zh) * 2019-11-21 2021-05-21 京东数字科技控股有限公司 手语识别方法及装置、计算机存储介质、电子设备
CN111859099B (zh) * 2019-12-05 2021-08-31 马上消费金融股份有限公司 基于强化学习的推荐方法、装置、终端以及存储介质
CN112929253B (zh) * 2019-12-05 2023-08-08 北京沃东天骏信息技术有限公司 一种虚拟形象交互方法和装置
CN112988956B (zh) * 2019-12-17 2024-07-12 北京搜狗科技发展有限公司 自动生成对话的方法及装置、信息推荐效果检测方法及装置
CN110929166B (zh) * 2019-12-27 2023-10-20 咪咕文化科技有限公司 一种内容推荐的方法、电子设备和存储介质
CN111159375A (zh) * 2019-12-27 2020-05-15 深圳前海微众银行股份有限公司 一种文本处理方法及装置
CN111159380B (zh) * 2019-12-31 2021-04-27 腾讯科技（深圳）有限公司 交互方法、装置、计算机设备及存储介质
CN111324710B (zh) * 2020-02-10 2024-01-12 深圳市医贝科技有限公司 一种基于虚拟人的在线调研方法、装置和终端设备
CN111309884A (zh) * 2020-02-14 2020-06-19 深圳壹账通智能科技有限公司 机器人对话方法、装置、介质及电子设备
CN111767721B (zh) * 2020-03-26 2025-01-10 北京沃东天骏信息技术有限公司 信息处理方法、装置及设备
CN111694939B (zh) * 2020-04-28 2023-09-19 平安科技（深圳）有限公司 智能调用机器人的方法、装置、设备及存储介质
CN111611365A (zh) * 2020-05-19 2020-09-01 上海鸿翼软件技术股份有限公司 一种对话系统的流程控制方法、装置、设备及存储介质
CN111723190A (zh) * 2020-06-29 2020-09-29 深圳壹账通智能科技有限公司 信息交互方法、电子设备及存储介质
CN112307168B (zh) * 2020-10-30 2023-11-07 康键信息技术（深圳）有限公司 基于人工智能的问诊会话处理方法、装置和计算机设备
CN112632246A (zh) * 2020-12-21 2021-04-09 平安普惠企业管理有限公司 基于深度学习的机器人对话方法、装置和计算机设备
CN113761176B (zh) * 2021-02-19 2025-03-18 北京沃东天骏信息技术有限公司 会话摘要提取方法、设备及存储介质
CN113157874B (zh) * 2021-02-20 2022-11-22 北京百度网讯科技有限公司 确定用户的意图的方法、装置、设备、介质和程序产品
CN113239146B (zh) * 2021-05-12 2023-07-28 平安科技（深圳）有限公司 应答分析方法、装置、设备及存储介质
CN114036279A (zh) * 2021-11-16 2022-02-11 平安普惠企业管理有限公司 机器人答复文本的确定方法、装置、设备及介质
CN114333806A (zh) * 2021-12-22 2022-04-12 深圳Tcl新技术有限公司 会话处理方法、装置、电子设备和计算机可读存储介质
CN114416937B (zh) * 2021-12-27 2023-06-20 北京百度网讯科技有限公司 人机交互方法、装置、设备、存储介质及计算机程序产品
CN114661882B (zh) * 2022-03-30 2024-09-13 浪潮通用软件有限公司 一种机器人聊天管理方法、设备及介质
CN115953883A (zh) * 2022-11-16 2023-04-11 深圳优地科技有限公司 配送系统的控制方法、机器人和存储介质
CN116541493A (zh) * 2023-04-07 2023-08-04 平安科技（深圳）有限公司 基于意图识别的交互应答方法、装置、设备、存储介质
CN116737883A (zh) * 2023-08-15 2023-09-12 科大讯飞股份有限公司 人机交互方法、装置、设备及存储介质
CN117171403A (zh) * 2023-09-06 2023-12-05 中国工商银行股份有限公司 数据处理方法、装置、计算机设备和存储介质
CN121166869A (zh) * 2023-09-07 2025-12-19 北京字跳网络技术有限公司 交互方法、装置及电子设备
CN116881429B (zh) * 2023-09-07 2023-12-01 四川蜀天信息技术有限公司 一种基于多租户的对话模型交互方法、装置及存储介质
CN119011514B (zh) * 2023-09-18 2025-11-25 北京字跳网络技术有限公司 信息处理的方法、装置、设备和存储介质
CN118377845B (zh) * 2024-05-06 2025-02-14 广东壹健康健康产业集团股份有限公司 基于向量计算的用户问答数据处理方法及系统
CN119030946A (zh) * 2024-08-18 2024-11-26 北京字跳网络技术有限公司 用于消息处理的方法、装置、设备和存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105068661A (zh) * 2015-09-07 2015-11-18 百度在线网络技术（北京）有限公司 基于人工智能的人机交互方法和系统
CN106997342A (zh) * 2017-03-27 2017-08-01 上海奔影网络科技有限公司 基于多轮交互的意图识别方法和装置
CN107918634A (zh) * 2017-06-27 2018-04-17 上海壹账通金融科技有限公司 智能问答方法、装置及计算机可读存储介质
US20180129646A1 (en) * 2011-03-31 2018-05-10 Microsoft Technology Licensing, Llc Augmented conversational understanding agent
CN109857848A (zh) * 2019-01-18 2019-06-07 深圳壹账通智能科技有限公司 交互内容生成方法、装置、计算机设备及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180129646A1 (en) * 2011-03-31 2018-05-10 Microsoft Technology Licensing, Llc Augmented conversational understanding agent
CN105068661A (zh) * 2015-09-07 2015-11-18 百度在线网络技术（北京）有限公司 基于人工智能的人机交互方法和系统
CN106997342A (zh) * 2017-03-27 2017-08-01 上海奔影网络科技有限公司 基于多轮交互的意图识别方法和装置
CN107918634A (zh) * 2017-06-27 2018-04-17 上海壹账通金融科技有限公司 智能问答方法、装置及计算机可读存储介质
CN109857848A (zh) * 2019-01-18 2019-06-07 深圳壹账通智能科技有限公司 交互内容生成方法、装置、计算机设备及存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115525740A (zh) * 2021-06-24 2022-12-27 Tcl科技集团股份有限公司 对话应答语句的生成方法、装置、电子设备及存储介质
CN117421398A (zh) * 2023-10-09 2024-01-19 北京百度网讯科技有限公司 人机交互方法、装置、设备以及存储介质

Also Published As

Publication number Publication date
CN109857848A (zh) 2019-06-07

Similar Documents

Publication Publication Date Title
WO2020147428A1 (zh) 交互内容生成方法、装置、计算机设备及存储介质
CN109829039B (zh) 智能聊天方法、装置、计算机设备及存储介质
US20230350929A1 (en) Method and system for generating intent responses through virtual agents
CN110765244B (zh) 获取应答话术的方法、装置、计算机设备及存储介质
CN107846350B (zh) 一种语境感知网络聊天的方法、计算机可读介质和系统
CN108090174B (zh) 一种基于系统功能语法的机器人应答方法及装置
US12456454B2 (en) System and method for a personalized dialogue system using knowledge-based entity services
WO2020177282A1 (zh) 一种机器对话方法、装置、计算机设备及存储介质
WO2021211200A1 (en) Natural language processing models for conversational computing
US20200027446A1 (en) Visualization interface for voice input
CN117520498A (zh) 基于虚拟数字人交互处理方法、系统、终端、设备及介质
CN120541310B (zh) 多源信息检索融合方法、装置、设备和可读存储介质
CN117033540A (zh) 报表生成方法、装置、电子设备和介质
US12530529B2 (en) Domain-specific named entity recognition via graph neural networks
CN116187346A (zh) 人机交互方法、装置、系统及介质
Guo et al. Who is answering whom? Finding “Reply-To” relations in group chats with deep bidirectional LSTM networks
CN117520544A (zh) 基于人工智能的信息鉴定方法、装置和计算机设备
CN120277199B (zh) 基于大模型的儿童教育知识边界管理方法、系统及设备
CN119783672B (zh) 语料扩展方法、设备及存储介质
CN115617992A (zh) 标签生成方法、装置、计算机可读存储介质及计算机设备
US20260064759A1 (en) Structured retrieval-augmented generation
CN119149699A (zh) 通话方法及其装置
CN118760733A (zh) 一种文本处理方法、装置、电子设备、计算机可读存储介质及计算机程序产品
CN114093447B (zh) 数据资产推荐方法、装置、计算机设备及存储介质
CN120892542B (zh) 一种智能对话方法、装置、设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19910032

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 02/11/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19910032

Country of ref document: EP

Kind code of ref document: A1