WO2019153522A1 - 智能交互方法、电子装置及存储介质 - Google Patents

智能交互方法、电子装置及存储介质 Download PDF

Info

Publication number
WO2019153522A1
WO2019153522A1 PCT/CN2018/084407 CN2018084407W WO2019153522A1 WO 2019153522 A1 WO2019153522 A1 WO 2019153522A1 CN 2018084407 W CN2018084407 W CN 2018084407W WO 2019153522 A1 WO2019153522 A1 WO 2019153522A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
intent
dialog
training
steps
Prior art date
Application number
PCT/CN2018/084407
Other languages
English (en)
French (fr)
Inventor
盛欢
吴焕明
邵诚
兰超
李发源
Original Assignee
卫盈联信息技术(深圳)有限公司
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 卫盈联信息技术(深圳)有限公司 filed Critical 卫盈联信息技术(深圳)有限公司
Publication of WO2019153522A1 publication Critical patent/WO2019153522A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36Creation of semantic tools, e.g. ontology or thesauri
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00Computing arrangements using knowledge-based models
    • G06N5/04Inference or reasoning models
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33Querying
    • G06F16/332Query formulation
    • G06F16/3329Natural language query formulation or dialogue systems
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Definitions

  • the present application relates to the field of natural language processing technologies, and in particular, to an intelligent interaction method, an electronic device, and a computer readable storage medium.
  • chat bots use deep learning to generate dialogues, prefer entertainment and chat, and have inaccurate semantic understanding.
  • traditional chat bots can only answer according to users' questions, and can not initiate communication to guide users to complete loan operations, interactive experience effects. difference.
  • the present application provides an intelligent interaction method, an electronic device, and a computer readable storage medium, the main purpose of which is to improve the accuracy of the intelligent question answering system, and at the same time, can guide the user to perform a process dialogue according to the input of the user, and actively guide the user to complete the process. Online business process.
  • an intelligent interaction method including:
  • Construction steps construct a dialog process template for different business processes, and train the intent to identify the classification model;
  • Processing step receiving a statement input by the user, performing a series of processing on the statement, including preprocessing, identifying the type of the sentence, and analyzing the emotion, and obtaining the emotional category of the user;
  • the identifying step identifying the user intent from the processed statement by using the pre-built dialog flow template and the trained intent recognition classification model;
  • Feedback step query the corresponding knowledge base according to the identified user intent, generate a reply in combination with the user emotion category, and feed back to the user.
  • the intent recognition classification model is a Convolutional Neural Network (CNN) model
  • CNN Convolutional Neural Network
  • Sorting steps Organizing training data, including obtaining user corpus and intent classification of user corpus;
  • Training steps training user corpus using skip-gram or continuous bag-of-words (CBOW) model to obtain low-dimensional word vectors;
  • CBOW continuous bag-of-words
  • Conversion step converting a low-dimensional word vector into a corresponding vector by looking up a word vector table
  • Extracting step setting the size of the window in the dimension of the feature matrix, converting the vector representation into a feature vector by convolution and pooling operations, and using the Hard Tanh function to extract the nonlinear feature of the feature vector, After the feature extraction of the preset number of times, the final feature of the user corpus is obtained;
  • Small class training steps For each subclass of the large class in the intent classification, the training methods using the above conversion steps, extraction steps and large class training steps are used for training.
  • the identifying step comprises: identifying whether the user is currently in the process dialogue, and if the user is in the process dialogue, parsing the conversation content to determine a corresponding dialog flow template, and guiding the dialog flow according to the pre-built dialog flow template, if the user Outside the process dialog, the pre-trained intent recognition classification model is invoked to identify the user's intent.
  • the identifying step further comprises: if the user has previously triggered the process dialogue and triggers the dialog again within the preset time, determining that the user is in the process dialogue, if the user first triggers the dialogue or exceeds the preset time again Trigger a conversation to identify if the user is currently in a process conversation.
  • the identifying step further comprises:
  • the intent recognition score is higher than or equal to the first threshold, determining that the user intention is classified into a business consultation class, and when the intention recognition score is lower than the first threshold and higher than the second threshold, determining that the user intention is classified into domain related in the chat class
  • the problem class when the intent recognition score is lower than or equal to the second threshold, judges that the user intends to classify into other problem classes in the chat class.
  • the pre-processing includes: word segmentation, part-of-speech tagging, Named Entities Recognition (NER), referencing disambiguation, and similar word expansion.
  • NER Named Entities Recognition
  • the present application further provides an electronic device, including: a memory, a processor, and a display, wherein the memory stores an intelligent interactive program, and the smart interactive program is executed by the processor, and the following steps can be implemented:
  • Construction steps construct a dialog process template for different business processes, and train the intent to identify the classification model;
  • Processing step receiving a statement input by the user, performing a series of processing on the statement, including preprocessing, identifying the type of the sentence, and analyzing the emotion, and obtaining the emotional category of the user;
  • the identifying step identifying the user intent from the processed statement by using the pre-built dialog flow template and the trained intent recognition classification model;
  • Feedback step query the corresponding knowledge base according to the identified user intent, generate a reply in combination with the user emotion category, and feed back to the user.
  • the present application further provides a computer readable storage medium, where the computer readable storage medium includes an intelligent interaction program, and when the intelligent interaction program is executed by a processor, the intelligent interaction as described above can be implemented. Any step in the method.
  • the intelligent interaction method, the electronic device and the computer readable storage medium proposed by the present application identify whether the user is in the dialogue process by performing a series of processing on the sentence input by the user. If the user is in the conversation process, the user dialogue process is guided according to the pre-built dialog flow template. If the user is outside the conversation process, the pre-trained model is used to identify the user's intent. Then, according to the user's intention, the corresponding knowledge base is queried, and the query result is fed back to the user in combination with the user's emotion, thereby improving the automation level of the user loan and improving the accuracy of the user sentence understanding.
  • FIG. 1 is a schematic diagram of a preferred embodiment of an electronic device of the present application.
  • FIG. 2 is a schematic diagram of a program module of the preferred embodiment of the intelligent interaction program of FIG. 1;
  • FIG. 3 is a schematic diagram of the function of the program module of Figure 2;
  • FIG. 5 is a flow chart of the present invention intended to identify a classification model training.
  • FIG. 1 is a schematic diagram of a preferred embodiment of an electronic device 1 of the present application.
  • the electronic device 1 may be a server, a smart phone, a tablet computer, a personal computer, a portable computer, and other electronic devices having computing functions.
  • the electronic device 1 includes a memory 11, a processor 12, a display 13, a network interface 14, and a communication bus 15.
  • the network interface 14 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).
  • Communication bus 15 is used to implement connection communication between these components.
  • the memory 11 includes at least one type of readable storage medium.
  • the at least one type of readable storage medium may be a non-volatile storage medium such as a flash memory, a hard disk, a multimedia card, a card type memory, or the like.
  • the memory 11 may be an internal storage unit of the electronic device 1, such as a hard disk of the electronic device 1.
  • the memory 11 may also be an external storage unit of the electronic device 1, such as a plug-in hard disk equipped on the electronic device 1, a smart memory card (SMC), and security. Digital (Secure Digital, SD) card, flash card (Flash Card), etc.
  • the memory 11 can be used to store application software and various types of data installed in the electronic device 1, such as the smart interaction program 10, a pre-built dialog flow template, and a trained intent recognition classification model.
  • the processor 12 in some embodiments, may be a Central Processing Unit (CPU), microprocessor or other data processing chip for running program code or processing data stored in the memory 11, such as executing an intelligent interactive program. 10 computer program code, training for intent to identify the classification model, and the like.
  • CPU Central Processing Unit
  • microprocessor or other data processing chip for running program code or processing data stored in the memory 11, such as executing an intelligent interactive program.
  • 10 computer program code, training for intent to identify the classification model, and the like.
  • Display 13 can be referred to as a display screen or display unit.
  • the display 13 can be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, and an Organic Light-Emitting Diode (OLED) touch device.
  • the display 13 is for displaying information processed in the electronic device 1 and a work interface for displaying visualization, such as displaying a user's sentence, replying, or asking a user's sentence.
  • FIG. 1 shows only the electronic device 1 having the components 11-15 and the intelligent interactive program 10, but it should be understood that not all illustrated components may be implemented, and more or fewer components may be implemented instead.
  • the electronic device 1 may further include a user interface
  • the user interface may include an input unit such as a keyboard, a voice output device such as an audio, a headphone, etc.
  • the user interface may further include a standard wired interface and a wireless interface. .
  • FIG. 2 it is a schematic diagram of a program module of the preferred embodiment of the intelligent interactive program in FIG.
  • the technical concept of the intelligent interaction method, program and device provided by the present application is illustrated by taking the loan business as an example, and other types of services are also applicable.
  • a module as referred to in this application refers to a series of computer program instructions that are capable of performing a particular function.
  • the smart interaction program 10 includes: a preprocessing module 110, a sentence type identification module 120, an sentiment analysis module 130, a dialog engine analysis module 140, a template engine module 150, an intent recognition training module 160, a service consulting module 170, The similarity module 180 and the dialog generation module 190 are retrieved.
  • module 110-190 The function of the module 110-190 is described below in conjunction with the functional diagram of the program module of FIG. 3:
  • the pre-processing module 110 is configured to pre-process the sentences input by the user, and the pre-processing includes: word segmentation, part-of-speech tagging, named entity identification, referencing disambiguation, and similar word expansion. Further, the word segmentation, part-of-speech tagging, and named entity recognition refer to a natural language processing tool trained by deep neural network for segmentation, part-of-speech tagging, and named entity recognition. The word segmentation refers to cutting a Chinese character sequence into a word sequence. The part-of-speech tagging refers to discriminating and labeling according to the part of speech of a word.
  • the part of speech includes: noun, adverb, adjective, verb, pronoun, etc., for example, NT stands for time noun, V stands for verb, NN stands for spoken noun, PU stands for calibration symbol, AD stands for adverbs, PN stands for pronouns, and so on.
  • the named entity identification refers to a named entity such as a person name, a place name, an organization name, and the like, and the named entity includes three major categories, such as an entity class, a time class, and a number class, and 7 subclasses, such as a person name, a place name, and an organization name. , time, date, currency, and percentage.
  • the reference disambiguation refers to eliminating the disambiguation of the personal pronouns, and using the Dependency Parsing (DP) to identify the interdependence between the components in the language unit to reveal its syntactic structure, to complete the disambiguation.
  • DP Dependency Parsing
  • the similar sentence expansion refers to the expansion of similar words by using Word2vec technology.
  • the sentence type identification module 120 is configured to combine the word segmentation, the result of the part-of-speech discrimination, and the dependency syntax analysis to identify the sentence type of the sentence input by the user.
  • the sentence types include: declarative sentences, imperative sentences, exclamatory sentences, and interrogative sentences, and the present application mainly includes interrogative sentences and declarative sentences.
  • the sentiment analysis module 130 is configured to combine the sentence type, input the processed sentence into a dictionary-based sentiment analysis method and a pre-trained deep learning-based sentiment classification model, and identify the user's emotional category.
  • the emotional categories include: anger, anxiety, anger, happiness, disappointment, surprise, curiosity, and the like.
  • the dictionary-based sentiment analysis method refers to constructing an sentiment analysis dictionary by a grammatically sensitive professional, and using a sentiment analysis dictionary according to the constructive sentiment analysis dictionary: a positive emotion dictionary, a negative emotion dictionary, and a neutral emotion dictionary.
  • the vocabulary expressing emotions is divided into three categories, and then the number of positive, negative and neutral emotion words in the sentence is compared, and the emotional category of the sentence is evaluated.
  • the deep learning-based sentiment classification model includes but is not limited to a Long Short-Term Memory (LSTM) model, a Support Vector Machine (SVM) model, a Random Forests (RF) model, and A pre-trained model in the Naive Bayesian Model (NBM) model.
  • LSTM Long Short-Term Memory
  • SVM Support Vector Machine
  • RF Random Forests
  • NBM Naive Bayesian Model
  • the dialog engine analysis module 140 is configured to identify whether the user is currently in the process dialog. If the user is in the process dialog, the template engine module 150 is invoked to parse the dialog content to determine a corresponding dialog flow template, and the dialog is guided according to the pre-built dialog flow template. The process, if the user is outside the process dialog, invokes the intent recognition training module 160 to identify the user intent.
  • the dialog process template includes, but is not limited to, a loan application process template, an audit process template, a lending process template, and a repayment process template.
  • the process template is maintained in a scripted language. However, it should be understood that if the user has previously triggered the process dialog, that is, the user intent has been initially identified, and the dialog is triggered again within the preset time, the default user is in the process dialog, and the template engine module 150 is invoked to resolve the dialog content to determine the corresponding Dialogue process template. If the user triggers the dialog for the first time or triggers the dialog after the preset time has elapsed, the dialog engine analysis module 140 is invoked to determine whether the user is currently in the process dialog. Assume that the preset time is 2 hours. If the user triggers the process dialog last time and triggers the dialog again after more than 2 hours, the dialog engine analysis module 140 is called to identify whether the user is currently in the process dialog.
  • the template engine module 150 is configured to search for a dialog process template that matches the current user intent, and guide the user according to the process set by the process template. For example, if the user's intention matches the loan application process template, the template engine module 150 guides the user to complete the data filling and data verification. If the user has the loan qualification, the amount of the loan that the user is required to be asked is guided. If the user does not qualify for the loan, the user is prompted to improve the information.
  • the intent recognition training module 160 is configured to identify the training of the classification model and input the user-processed sentence into the intent recognition classification model to identify the user's intention.
  • the user intention classification is divided into two categories, including: loan consultation class and chat class. Further, each major category is further divided into several small categories for precise positioning of user questions.
  • the loan consultation category includes: application qualification consultation, application quota consultation, mortgage consultation, repayment method consultation, audit duration consultation, repayment plan consultation...
  • the intent recognition training module 160 inputs the processed statement into the intent classification model, An intent recognition score is output for each category, and the category corresponding to the maximum value is the user's intention.
  • the intent recognition score when the intent recognition score is higher than or equal to the first threshold, it is determined that the user intention is classified into a loan consultation class, and the business consultation module 170 is invoked to query the answer from the domain knowledge base.
  • the intent recognition score is lower than the first threshold and higher than the second threshold, it is determined that the user intends to classify into a domain related problem class in the chat class, and the search similarity module 180 is invoked to query the answer from the domain related knowledge base.
  • the intent recognition score is lower than or equal to the second threshold, it is determined that the user intends to classify into other problem classes in the chat class, and the search similarity module 180 is invoked to query the answer from the gossip question answering system.
  • the intent recognition classification model is pre-trained by the CNN model, and the training steps for intent to identify the classification model are as follows:
  • Sorting steps Organizing training data, including obtaining user corpus and intent classification of user corpus;
  • Training steps training user corpus using skip-gram or CBOW model to obtain low-dimensional word vectors
  • Conversion step converting a low-dimensional word vector into a corresponding vector by looking up a word vector table
  • Extracting step setting the size of the window in the dimension of the feature matrix, converting the vector representation into a feature vector by convolution and pooling operations, and using the Hard Tanh function to extract the nonlinear feature of the feature vector, After the feature extraction of the preset number of times, the final feature of the user corpus is obtained;
  • Small class training steps For each subclass of the large class in the intent classification, the training methods using the above conversion steps, extraction steps and large class training steps are used for training.
  • the business consulting module 170 is configured to query an answer from the domain knowledge base according to the user's intention.
  • the self-domain knowledge base can solve the problem of high degree of relevance in the business field. For example, to solve the problem of high relevance of users in the field of loans. Assuming that the user asks about the loan interest rate question, after identifying the user's intent, the query gets the answer to the question from the domain knowledge base.
  • the answers from the domain knowledge base can be maintained by customer service personnel. When a user's problem belongs to the loan consultation class and each small category under the loan consultation category has a lower intent recognition score, the customer service staff can determine the problem as a new problem and add the problem to the domain knowledge base. The answer makes the semantic understanding of the system more and more precise.
  • the search similarity module 180 is configured to query an answer from the domain related knowledge base and the gossip question answering system according to the user's intention. Wherein, when the user intends to classify the domain related problem class in the chat class, the answer is obtained from the domain related knowledge base. When the user intent type is another question class in the chat class, the answer is queried from the gossip question answering system, which is used to solve the completely unrelated question.
  • the domain-related knowledge base is built based on crawler technology to solve domain-related consulting problems. For example, in order to solve the financial unpopular problems of some users' questions, crawling technology is used to crawl some knowledge related to the construction of financial fields from the Internet. Library. The knowledge base is searched by means of retrieval. If the intent recognition score corresponding to all the small categories of the knowledge base is too low, the problem raised by the user is compared with the problem in the domain-related knowledge base, and the answer corresponding to the question with the highest similarity score in the domain-related knowledge base is used to answer the user. . However, it should be understood that there are many problems in the domain-related knowledge base and their corresponding answers.
  • the domain-related knowledge base uses the pre-trained Locality-Sensitive Hashing Forest (LSH Forest) model. Divide all the data, reduce the number of points per search and calculation to an acceptable range, and then establish multiple LSH Forests, taking the combined results of the forest as the final result.
  • LSH Forest Locality-Sensitive Hashing Forest
  • the chat question and answer system refers to a form in which a chat conversation is organized into a question and answer pair. Further, when the question asked by the user is similar to the problem built in the chat question and answer system, the crawler can use the crawler to crawl the relevant answer to reply, and add the form of the answer and the question form to the chat system. in. It should be understood that the gossip question and answer system also uses the LSH Forest model to divide all question and answer pairs.
  • the dialog generating module 190 is configured to generate a reply according to the user's emotional category, the user's intention, and the query result. It should be understood that the responses are in the same intent, and different emotional responses may be set for different emotions of the user. For example, when the user's emotion is detected as anxiety, a soothing emotion and tone can be added to the reply.
  • FIG. 4 it is a flowchart of a preferred embodiment of the intelligent interaction method of the present application.
  • Step S10-Step S40 when the processor 12 executes the computer program of the intelligent interactive program 10 stored in the memory 11, implementing the intelligent interaction method includes: Step S10-Step S40:
  • step S10 the dialog process template and the training intention identification classification model of different business processes are pre-built.
  • the dialog process template includes, but is not limited to, a loan application process template, an audit process template, a lending process template, and a repayment process template.
  • the intent recognition classification model is pre-trained by the CNN model, as shown in FIG. 5, which is a flowchart of the present invention intended to identify the classification model training, and the training steps of the intent to identify the classification model are as follows:
  • step S11 the training data is organized, including obtaining the user corpus and intent classification of the user corpus. For example, get 200,000 user statements that cover all of the small categories of user intent types, and label large categories of all statements and small categories under large categories.
  • the large categories of the user intent types are classified into three categories, including: a loan consultation class and a chat class, wherein the chat class includes: a domain related question class and other problem classes.
  • step S12 the user corpus is trained by using the skip-gram or CBOW model to obtain a low-dimensional word vector.
  • the CBOW model is used to train the statements to get the low-dimensional word vector for each word in each statement.
  • step S13 the low-dimensional word vector is converted into a corresponding vector by looking up the word vector table.
  • the word vector table is pre-built, converting different two or more low-dimensional word vectors into corresponding vectors in the table.
  • Step S14 setting a size of the window in a dimension of the feature matrix, converting the vector representation into a feature vector by a convolution and a pooling operation, and performing a nonlinear feature extraction using the Hard Tanh function on the feature vector.
  • the final characteristics of the user corpus are obtained.
  • the setting window is 128-dimensional
  • the vector converted from the word vector table is converted into a feature vector
  • the Hard Tanh function is used to extract the nonlinear features to obtain the final feature.
  • step S15 the gradient descent algorithm is used to iteratively, and the supervised training is performed on the large classes in the intent classification.
  • the gradient descent algorithm refers to a matrix method. After iterative training, a model that identifies a large category of user intent types is generated.
  • each small category of the large category in the intent type is trained using the training methods of the above steps S13, S14, and S15 to generate a model that can identify the small category of the user's intention type.
  • Step S20 is a program module of the intelligent interaction method of the present application and a functional diagram thereof.
  • the user 2 input statement is received, and corresponding processing is performed on the statement, including preprocessing, sentence type recognition, and sentiment analysis, to obtain the emotion category of the user 2.
  • the pre-processing refers to the pre-processing module 110 performing word segmentation, part-of-speech tagging, named entity identification, referencing disambiguation, and similar term expansion processing on the sentence input by the user 2.
  • the sentence type identification module 120 recognizes the sentence type of the sentence input by the user 2 in combination with the result of the word segmentation, the part of speech recognition, and the DP. For example, based on the analysis of words such as "how long" in the input sentence of the user 2, the sentence is recognized as a question.
  • the sentiment analysis module 130 combines the sentence type, inputs the processed sentence into the dictionary-based sentiment analysis mode and the pre-trained deep learning-based sentiment classification model, and identifies the emotion category of the user 2. For example, combined with the sentence type, the words "how long", “one week” and other words identify the user's emotional polarity as negative, and at the same time analyze the main emotions of user 2 at this time as “anger” and "anxiety.”
  • step S30 the user intent is identified from the processed statement by using the pre-built dialog flow template and the trained intent recognition classification model.
  • the dialog engine analysis module 140 identifies whether the user 2 is currently in the dialog flow. If the user 2 is currently in the dialog flow, the template engine module 160 is invoked to parse the statement content, determine the corresponding dialog flow template, and guide the user according to the dialog flow template. Process dialogue. If the user 2 has triggered the dialog process before, and triggers the dialog again within a preset time, such as within 2 hours, the user 2 is in the current dialog flow, and the template engine module 160 is called to parse the sentence content to determine the corresponding dialog. Process template.
  • the dialog engine analysis module 140 re-identifies whether the user is currently in the process dialog.
  • the loan consultation category includes, but is not limited to, application qualification consultation, application quota consultation, mortgage consultation, repayment consultation, audit duration consultation, and repayment plan consultation.
  • the loan consultation class is a large category of user intent types, and the application qualification consultation class, the application quota consultation class, the mortgage consultation class, the repayment mode consultation class, the audit time consultation class, and the repayment plan consultation class are all user intention types. Small category.
  • the intent recognition training module 150 processes the sentence of the user 2 and inputs the trained intent recognition classification model, and recognizes that the intention of the user 2 is the audit duration consultation class, and needs to know the loan review duration.
  • step S40 the corresponding knowledge base is queried according to the identified user intention, and the query result is combined with the user emotion category to generate a reply, which is fed back to the user 2.
  • the business consultation module 170 queries the domain knowledge base 31, and the loan review period is two weeks, then the dialog generation module 190 combines the user. 2 "angry" and “anxiety” sentiment replied to User 2: “Hello, Dear User! Don't worry, your loan will be reviewed after a week, thank you!
  • the statement entered by the user 2 is: "I want to loan 20,000 yuan.”
  • the result of the word segmentation and part-of-speech tagging is: "I / PN want / VV loan / V20000 / M yuan / RN. / PU”
  • the named entity recognizes the currency named entity "20,000 yuan” in the statement.
  • the sentence type identification module 120 recognizes the sentence type of the sentence input by the user 2 in combination with the word segmentation, the result of the part-of-speech discrimination, and the dependency syntax analysis. For example, based on the analysis of words such as "credit" in the input sentence of the user 2, the statement is recognized as a declarative sentence.
  • the sentiment analysis module 130 combines the sentence type, inputs the processed sentence into the dictionary-based sentiment analysis mode and the pre-trained deep learning-based sentiment classification model, and identifies the emotion category of the user 2. For example, combined with the sentence type, the user's emotional polarity is negative through words such as “credit”, and the main emotion of user 2 is “anxiety” at this time.
  • the dialog engine analysis module 140 identifies that the user 2 is currently in the dialog flow, and invokes the template engine module 150 to find a dialog flow template that matches the current user's intention as a loan application process template, and guides the user to complete the data filling and data verification if the user has a loan. Qualification, inform the user of the loan amount issuance time, repayment date and repayment amount. If the user does not qualify for the loan, the user is prompted to improve the information.
  • the intelligent interaction method proposed by the foregoing embodiment processes the statement input by the user, analyzes the type of the statement, the emotion of the user, and then determines whether the user is in the process dialogue, and the user is in the process, and determines the dialog flow template to guide the user to the loan. If the user is not in the process, perform related operations such as loan consultation and chat, according to different intentions of the user, generate different responses for the user in combination with user emotions, interact with the user, improve automation level and user experience, and improve semantics. Understand the accuracy.
  • the embodiment of the present application further provides a computer readable storage medium, where the computer readable storage medium includes an intelligent interaction program 10, and when the intelligent interaction program 10 is executed by the processor, the following operations are implemented:
  • Construction steps construct a dialog process template for different business processes, and train the intent to identify the classification model;
  • Processing step receiving a statement input by the user, performing a series of processing on the statement, including preprocessing, identifying the type of the sentence, and analyzing the emotion, and obtaining the emotional category of the user;
  • the identifying step identifying the user intent from the processed statement by using the pre-built dialog flow template and the trained intent recognition classification model;
  • Feedback step query the corresponding knowledge base according to the identified user intent, generate a reply in combination with the user emotion category, and feed back to the user.
  • the intent recognition classification model is a CNN model
  • the training steps of the intent to identify the classification model are as follows:
  • Sorting steps Organizing training data, including obtaining user corpus and intent classification of user corpus;
  • Training steps training user corpus using skip-gram or CBOW model to obtain low-dimensional word vectors
  • Conversion step converting a low-dimensional word vector into a corresponding vector by looking up a word vector table
  • Extracting step setting the size of the window in the dimension of the feature matrix, converting the vector representation into a feature vector by convolution and pooling operations, and using the Hard Tanh function to extract the nonlinear feature of the feature vector, After the feature extraction of the preset number of times, the final feature of the user corpus is obtained;
  • Small class training steps For each subclass of the large class in the intent classification, the training methods using the above conversion steps, extraction steps and large class training steps are used for training.
  • the identifying step comprises: identifying whether the user is currently in the process dialogue, and if the user is in the process dialogue, parsing the conversation content to determine a corresponding dialog flow template, and guiding the dialog flow according to the pre-built dialog flow template, if the user Outside the process dialog, the pre-trained intent recognition classification model is invoked to identify the user's intent.
  • the identifying step further comprises: if the user has previously triggered the process dialogue and triggers the dialog again within the preset time, determining that the user is in the process dialogue, if the user first triggers the dialogue or exceeds the preset time again Trigger a conversation to identify if the user is currently in the process conversation.
  • the identifying step further comprises:
  • the intent recognition score is higher than or equal to the first threshold, determining that the user intention is classified into a business consultation class, and when the intention recognition score is lower than the first threshold and higher than the second threshold, determining that the user intention is classified into domain related in the chat class
  • the problem class when the intent recognition score is lower than or equal to the second threshold, judges that the user intends to classify into other problem classes in the chat class.
  • the pre-processing includes: word segmentation, part-of-speech tagging, NER, referencing disambiguation, and similar word expansion.
  • the technical solution of the present application which is essential or contributes to the prior art, may be embodied in the form of a software product stored in a storage medium (such as ROM/RAM as described above). , a disk, an optical disk, including a number of instructions for causing a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to perform the methods described in the various embodiments of the present application.
  • a terminal device which may be a mobile phone, a computer, a server, or a network device, etc.

Abstract

本申请提供了一种智能交互方法、电子装置及存储介质,该方法预先构建不同业务流程的对话流程模板,并训练意图识别分类模型。接收到用户输入的语句后,对语句执行一系列处理,包括预处理、识别语句类型及情感分析,识别用户的情绪类别,并利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的语句中识别用户意图。之后,根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户。利用本申请,不仅能够提高智能问答系统的准确率,还可以根据用户的输入引导用户执行流程对话,主动引导用户完成线上业务流程。

Description

智能交互方法、电子装置及存储介质
优先权申明
本申请要求于2018年02月09日提交中国专利局、申请号为201810134194.2,名称为“智能交互方法、电子装置及存储介质”的中国专利申请的优先权,该中国专利申请的整体内容以参考的方式结合本申请中。
技术领域
本申请涉及自然语言处理技术领域,尤其涉及一种智能交互方法、电子装置及计算机可读存储介质。
背景技术
近年来,随着人工智能应用领域的不断扩大,各领域相关公司都相续开发自己的聊天机器人。在现有的贷款申请流程中,由于牵涉到钱财、个人信息等一系列安全问题,用户在贷款过程中可能需要咨询大量的问题。传统的聊天机器人使用深度学习生成对话,偏向于娱乐闲聊,对语义理解不精准,同时传统的聊天机器人只能根据用户的提问进行回答,不能够主动发起交流引导用户完成贷款操作,交互式体验效果差。
发明内容
鉴于以上内容,本申请提供一种智能交互方法、电子装置及计算机可读存储介质,其主要目的在于提高智能问答系统的准确率,同时能够根据用户的输入引导用户执行流程对话,主动引导用户完成线上业务流程。
为实现上述目的,本申请提供一种智能交互方法,该方法包括:
构建步骤:构建不同业务流程的对话流程模板,并训练意图识别分类模型;
处理步骤:接收用户输入的语句,对语句执行一系列处理,包括预处理、识别语句类型以及情感分析,得到用户的情绪类别;
识别步骤:利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的语句中识别用户意图;
反馈步骤:根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户。
优选地,所述意图识别分类模型为卷积神经网络(Convolutional Neural Network,CNN)模型,所述意图识别分类模型的训练步骤如下:
整理步骤:整理训练数据,包括获取用户语料以及标注用户语料的意图分类;
训练步骤:利用skip-gram或者连续词袋(Continuous Bag-of-Words,CBOW)模型训练用户语料,得到低维度的字向量;
转换步骤:通过查找字向量表的方式,将低维度的字向量转换成相应的向量;
抽取步骤:在特征矩阵的维度上,设定窗口的大小,通过卷积、池化操作将所述向量表示式转换成特征向量,对所述特征向量使用Hard Tanh函数进行非线性特征的抽取,经过预设次数的特征抽取之后,获取用户语料的最终特征;
大类训练步骤:利用梯度下降算法迭代,对意图分类中的大类进行有监督的训练;
小类训练步骤:对意图分类中大类的每个小类,使用上述转换步骤、抽取步骤及大类训练步骤的训练方式进行训练。
优选地,所述识别步骤包括:识别用户当前是否在流程对话中,若用户处于流程对话中,则解析对话内容以确定对应的对话流程模板,根据预先构建的对话流程模板引导对话流程,若用户处于流程对话外,则调用预先训练好的意图识别分类模型识别用户意图。
优选地,所述识别步骤还包括:若用户之前已触发流程对话,并且在预设时间内再次触发对话,则判断用户处于流程对话中,若用户第一次触发对话或者超过预设时间后再次触发对话,则识别用户当前是否处于流程对话中。
优选地,所述识别步骤还包括:
将所述语句输入意图识别分类模型,得到该语句的意图识别分数;
当意图识别分数高于或等于第一阈值时,判断用户意图分类为业务咨询类,当意图识别分数低于第一阈值且高于第二阈值时,判断用户意图分类为闲聊类中的领域相关问题类,当意图识别分数低于或等于第二阈值时,判断 用户意图分类为闲聊类中的其他问题类。
优选地,所述预处理包括:分词、词性标注、命名实体识别(Named Entities Recognition,NER)、指代消歧,以及相似词语扩展。
此外,本申请还提供一种电子装置,该电子装置包括:存储器、处理器及显示器,所述存储器上存储智能交互程序,所述智能交互程序被所述处理器执行,可实现如下步骤:
构建步骤:构建不同业务流程的对话流程模板,并训练意图识别分类模型;
处理步骤:接收用户输入的语句,对语句执行一系列处理,包括预处理、识别语句类型以及情感分析,得到用户的情绪类别;
识别步骤:利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的语句中识别用户意图;
反馈步骤:根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户。
此外,为实现上述目的,本申请还提供一种计算机可读存储介质,所述计算机可读存储介质中包括智能交互程序,所述智能交互程序被处理器执行时,可实现如上所述智能交互方法中的任意步骤。
本申请提出的智能交互方法、电子装置及计算机可读存储介质,通过对用户输入的语句进行一系列处理,识别用户是否处于对话流程中。若用户处于对话流程中,则根据预先构建的对话流程模板引导用户对话流程。若用户处于对话流程之外,则利用预先训练好的模型识别用户意图。之后根据用户意图查询相应的知识库,将查询结果结合用户情绪反馈给用户,从而提升用户贷款的自动化水平,提高用户语句理解的准确度。
附图说明
图1为本申请电子装置较佳实施例的示意图;
图2为图1中智能交互程序较佳实施例的程序模块示意图;
图3为图2中程序模块的功能示意图;
图4为本申请智能交互方法较佳实施例的流程图;
图5为本申请意图识别分类模型训练的流程图。
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
具体实施方式
应当理解,此处所描述的具体实施例仅用以解释本申请,并不用于限定本申请。
如图1所示,是本申请电子装置1较佳实施例的示意图。
在本实施例中,电子装置1可以是服务器、智能手机、平板电脑、个人电脑、便携计算机以及其它具有运算功能的电子设备。
该电子装置1包括:存储器11、处理器12、显示器13、网络接口14及通信总线15。其中,网络接口14可选地可以包括标准的有线接口、无线接口(如WI-FI接口)。通信总线15用于实现这些组件之间的连接通信。
存储器11至少包括一种类型的可读存储介质。所述至少一种类型的可读存储介质可为如闪存、硬盘、多媒体卡、卡型存储器等的非易失性存储介质。在一些实施例中,所述存储器11可以是所述电子装置1的内部存储单元,例如该电子装置1的硬盘。在另一些实施例中,所述存储器11也可以是所述电子装置1的外部存储单元,例如所述电子装置1上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。
在本实施例中,所述存储器11可以用于存储安装于所述电子装置1的应用软件及各类数据,例如智能交互程序10、预先构建的对话流程模板及训练好的意图识别分类模型。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU),微处理器或其它数据处理芯片,用于运行存储器11中存储的程序代码或处理数据,例如执行智能交互程序10的计算机程序代码、意图识别分类模型的训练等。
显示器13可以称为显示屏或显示单元。在一些实施例中显示器13可以 是LED显示器、液晶显示器、触控式液晶显示器以及有机发光二极管(Organic Light-Emitting Diode,OLED)触摸器等。显示器13用于显示在电子装置1中处理的信息以及用于显示可视化的工作界面,例如显示用户的语句、回复或提问用户的语句。
图1仅示出了具有组件11-15以及智能交互程序10的电子装置1,但是应理解的是,并不要求实施所有示出的组件,可以替代的实施更多或者更少的组件。
可选地,该电子装置1还可以包括用户接口,用户接口可以包括输入单元比如键盘(Keyboard)、语音输出装置比如音响、耳机等,可选地用户接口还可以包括标准的有线接口、无线接口。
如图2所示,是图1中智能交互程序较佳实施例的程序模块示意图。
在实施方式中,以贷款业务为例阐述本申请提供的智能交互方法、程序及装置的技术构思,其它类型的业务同样适用。
本申请所称的模块是指能够完成特定功能的一系列计算机程序指令段。
在本实施例中,智能交互程序10包括:预处理模块110、语句类型识别模块120、情感分析模块130、对话引擎分析模块140、模板引擎模块150、意图识别训练模块160、业务咨询模块170、检索相似度模块180及对话生成模块190。
以下结合图3的程序模块的功能示意图说明模块110-190的功能:
预处理模块110,用于对用户输入的语句进行预处理,所述预处理包括:分词、词性标注、命名实体识别、指代消歧、相似词语扩展。进一步地,所述分词、词性标注、命名实体识别是指使用深度神经网络训练的自然语言处理工具进行分词、词性标注、命名实体识别。所述分词是指将汉字序列切分成词序列。所述词性标注是指根据词语的词性进行判别及标注,词性包括:名词、副词、形容词、动词、代词等,例如,NT代表时间名词、V代表动词、NN代表口头名词、PU代表标定符号、AD代表副词、PN代表代词等。所述命名实体识别是指识别语句中人名、地名、组织机构名等命名实体,命名实体包括3大类,如实体类、时间类及数字类,和7小类,如人名、地名、机构名、时间、日期、货币及百分比。所述指代消歧是指消除人称代词的指代歧义,通过使用依存句法分析(Dependency Parsing,DP)识别语言单位内成 分之间的依存关系揭示其句法结构,来完成指代消歧。所述相似语句扩展是指利用Word2vec技术进行相似词语的扩展。
语句类型识别模块120,用于结合分词、词性判别的结果以及依存句法分析识别出用户输入的语句的句子类型。所述句子类型包括:陈述句、祈使句、感叹句、疑问句,本申请中主要包括疑问句和陈述句。
情感分析模块130,用于结合句子类型,将处理后的语句输入基于词典的情感分析方式和预先训练好的基于深度学习的情感分类模型,识别出用户的情绪类别。所述情绪类别包括:愤怒、焦虑、生气、开心、失望、惊讶、好奇等。所述基于词典的情感分析方式是指通过具有语法敏感性的专业人士构建情感分析词典,根据构建的情感分析词典:正性情感词典、负性情感词典及中性情感词典,将某语句中用于表达情感的词汇分为三个类别,然后对比语句中正性、负性及中性情感词的个数,评估语句的情绪类别。所述基于深度学习的情感分类模型包括但不限于长短期记忆网络(Long Short-Term Memory,LSTM)模型、支持向量机(Support Vector Machine,SVM)模型、随机森林(Random Forests,RF)模型及朴素贝叶斯(Naive Bayesian Model,NBM)模型中一种预先训练好的模型。该模型是由已人工识别正、负及中性的文本通过机器学习等方式训练而成,在此不再赘述。
对话引擎分析模块140,用于识别用户当前是否在流程对话中,若用户处于流程对话中,则调用模板引擎模块150解析对话内容以确定对应的对话流程模板,根据预先构建的对话流程模板引导对话流程,若用户处于流程对话之外,则调用意图识别训练模块160识别用户意图。
所述对话流程模板包括但不限于贷款申请流程模板、审核流程模板、放款流程模板、还款流程模板。所述流程模板是通过脚本式的语言进行维护的。但应理解的是,若用户之前已触发流程对话,即已初步识别用户意图,并且在预设时间内再次触发对话,则默认用户处于流程对话中,调用模板引擎模块150解析对话内容以确定对应的对话流程模板。若用户第一次触发对话或者超过预设时间后再触发对话,则调用对话引擎分析模块140判断用户当前是否在流程对话中。假设,预设时间为2小时,若用户上一次触发流程对话,超过2个小时之后再次触发对话,则调用对话引擎分析模块140识别该用户当前是否在流程对话中。
模板引擎模块150,用于查找与当前用户意图相匹配的对话流程模板,依据流程模板设定的流程引导用户。例如,用户意图与贷款申请流程模板相匹配,则模板引擎模块150引导用户完成资料填写、资料校验,如果用户具有贷款资格,则引导提问用户需要贷款的金额。如果用户不具备贷款资格,提示用户完善资料。
意图识别训练模块160,用于意图识别分类模型的训练以及将用户处理后的语句输入意图识别分类模型中,识别用户的意图。本实施例中,用户意图分类分为两大类,包括:贷款咨询类及闲聊类。进一步地,每一大类别又分为若干个小类别,用于对用户问题进行精确定位。例如,贷款咨询类包括:申请资格咨询、申请额度咨询、抵押咨询、还款方式咨询、审核时长咨询、还款计划咨询……意图识别训练模块160将处理后的语句输入意图识分类模型后,针对每个类别会输出一个意图识别分数,取最大值对应的类别为用户的意图。其中,当意图识别分数高于或等于第一阈值时,判断用户意图分类为贷款咨询类,调用业务咨询模块170从自领域知识库中查询答案。当意图识别分数低于第一阈值且高于第二阈值时,判断用户意图分类为闲聊类中的领域相关问题类,调用检索相似度模块180从领域相关知识库中查询答案。当意图识别分数低于或等于第二阈值时,判断用户意图分类为闲聊类中的其他问题类,调用检索相似度模块180从闲聊问答系统中查询答案。假设,第一阈值为60分,第二阈值为30分,当意图识别分数为46分时,判断用户意图分类为领域相关问题类。所述意图识别分类模型是CNN模型预先训练得到的,意图识别分类模型的训练步骤如下:
整理步骤:整理训练数据,包括获取用户语料以及标注用户语料的意图分类;
训练步骤:利用skip-gram或者CBOW模型训练用户语料,得到低维度的字向量;
转换步骤:通过查找字向量表的方式,将低维度的字向量转换成相应的向量;
抽取步骤:在特征矩阵的维度上,设定窗口的大小,通过卷积、池化操作将所述向量表示式转换成特征向量,对所述特征向量使用Hard Tanh函数进行非线性特征的抽取,经过预设次数的特征抽取之后,获取用户语料的最终 特征;
大类训练步骤:利用梯度下降算法迭代,对意图分类中的大类进行有监督的训练;
小类训练步骤:对意图分类中大类的每个小类,使用上述转换步骤、抽取步骤及大类训练步骤的训练方式进行训练。
业务咨询模块170,用于根据用户的意图,从自领域知识库中查询答案。所述自领域知识库可以解决业务领域相关程度高的问题。例如,解决用户在贷款领域相关程度高的问题。假设用户询问关于贷款利率问题时,识别出该用户的意图后,查询自领域知识库得到该问题的答案。所述自领域知识库的答案可以由客服人员进行维护。当某用户的问题属于贷款咨询类且贷款咨询类下的每个小类别对应的意图识别分数均较低时,客服人员可以确定该问题为新问题,并在自领域知识库中增加该问题的答案,从而使得系统的语义理解越来越精确。
检索相似度模块180,用于根据用户意图,从领域相关知识库及闲聊问答系统查询答案。其中,当用户意图分类为闲聊类中的领域相关问题类时,从领域相关知识库查询答案。当用户意图类型为闲聊类中的其他问题类时,从闲聊问答系统查询答案,所述闲聊问答系统用于解决完全无关问题。
所述领域相关知识库是基于爬虫技术构建的,用以解决领域相关咨询问题,例如,为了解决一些用户提问的金融类冷门问题,利用爬虫技术从互联网爬取部分金融领域相关知识构建领域相关知识库。该知识库采用检索的方式进行检索。若该知识库的所有小类别对应的意图识别分数过低,则将用户提出的问题与领域相关知识库中的问题进行对比,采纳领域相关知识库中相似度分数最高的问题对应的答案答复用户。但应理解的是,领域相关知识库中的问题及其对应的答案数目较多,因此,领域相关知识库中使用预先训练好的局部敏感随机投影森林(Locality-Sensitive Hashing Forest,LSH Forest)模型对所有的数据进行划分,将每次搜索与计算的点的数目减小到一个可接受的范围,然后建立多个LSH Forest,将森林的综合结果作为最终的结果。
所述闲聊问答系统是指将闲聊对话整理成问答对的形式构建的。进一步的,当用户提问的问题与闲聊问答系统中构建的问题相似度较低,可以使用爬虫从互联网中爬取相关的答案进行回复,并将该答案与问题形成问答对的 形式添加进闲聊系统中。应理解的是,闲聊问答系统同样使用LSH Forest模型对所有的问答对进行划分。
对话生成模块190,用于根据用户的情绪类别、用户意图及查询结果生成回复。应理解的是,所述回复在相同的意图中,针对用户不同的情绪可以设置不同的情感回复。例如,检测到用户的情绪为焦虑时,可以在回复中加入安抚性的情感及语气。
如图4所示,是本申请智能交互方法较佳实施例的流程图。
在本实施例中,处理器12执行存储器11中存储的智能交互程序10的计算机程序时实现智能交互方法包括:步骤S10-步骤S40:
步骤S10,预先构建不同业务流程的对话流程模板及训练意图识别分类模型。所述对话流程模板包括但不限于贷款申请流程模板、审核流程模板、放款流程模板、还款流程模板。所述意图识别分类模型是CNN模型预先训练得到的,如图5所示,是本申请意图识别分类模型训练的流程图,意图识别分类模型的训练步骤如下:
步骤S11,整理训练数据,包括获取用户语料以及标注用户语料的意图分类。例如,获取20万个用户语句,该语句涉及用户意图类型所有的小类别,并标注所有语句的大类别及大类别之下的小类别。所述用户意图类型的大类别分为三大类,包括:贷款咨询类及闲聊类,其中,闲聊类又包括:领域相关问题类和其他问题类。
步骤S12,利用skip-gram或者CBOW模型训练用户语料,得到低维度的字向量。例如,使用CBOW模型对语句进行训练,得到每个语句中每个字的低维度字向量。
步骤S13,通过查找字向量表的方式,将低维度的字向量转换成相应的向量。所述的字向量表是预先构建完成的,将不同的两个或多个低维度字向量转换成表格中对应的向量。
步骤S14,在特征矩阵的维度上,设定窗口的大小,通过卷积、池化操作将所述向量表示式转换成特征向量,对所述特征向量使用Hard Tanh函数进行非线性特征的抽取,经过预设次数的特征抽取之后,获取用户语料的最终特征。例如设定窗口为128维,将字向量表转成的向量转换为特征向量,并用Hard  Tanh函数进行非线性特征的抽取得到最终特征。
步骤S15,利用梯度下降算法迭代,对意图分类中的大类进行有监督的训练。其中,所述梯度下降算法是指矩阵法。进行迭代训练后,生成能够识别用户意图类型的大类别的模型。
步骤S16,对意图类型中大类别的每个小类别,使用上述步骤S13、S14、S15的训练方式进行训练,生成能够识别用户意图类型的小类别的模型。
步骤S20,如图5所示,是本申请智能交互方法的程序模块及其功能示意图。接收用户2输入语句,对该语句执行相应的处理,包括预处理、语句类型识别及情感分析,得到用户2的情绪类别。所述预处理是指预处理模块110对用户2输入的语句进行分词、词性标注、命名实体识别、指代消歧、相似词语扩展处理。假设,用户2输入的语句为:“我提交的借款一星期了,还要多久能审核下来。”,分词及词性标注后的结果为:“我/PN提交/V的/U借款/NN一星期/NT了/U,/PU还/AD要/V多久/AD能/V审核/V下来/V。/PU”,命名实体识别出该语句中的“一星期”。语句类型识别模块120结合分词、词性判别的结果以及DP识别出用户2输入的语句的句子类型。例如,根据该用户2输入语句中的“多久”等词语的分析,识别出该语句为疑问句。情感分析模块130结合句子类型,将处理后的语句输入基于词典的情感分析方式和预先训练好的基于深度学习的情感分类模型,识别出用户2的情绪类别。例如,结合句子类型,通过“多久”、“一星期”等词语识别出用户情绪极性为负性,同时分析出此时用户2的主要情绪为“愤怒”和“焦虑”。
步骤S30,利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的该语句中识别用户意图。对话引擎分析模块140识别用户2当前是否在对话流程中,若用户2当前在对话流程中,则调用模板引擎模块160解析语句内容,确定对应的对话流程模板,并根据该对话流程模板引导用户2的流程对话。其中,若用户2之前已触发过对话流程,并在预设时间,如2个小时内再次触发对话,则默认该用户2当前对话流程中,调用模板引擎模块160解析语句内容,确定对应的对话流程模板。若用户2首次或超过预设时间,如2个小时后输入语句触发对话流程,则对话引擎分析模块140重新识别用户当前是否在流程对话中。假设,贷款咨询类包括但不限于申请资格咨询类、申请额度咨询类、抵押咨询类、还款方式咨询类、审核时长咨询类及还款计划咨 询类。所述贷款咨询类为用户意图类型的大类别,所述申请资格咨询类、申请额度咨询类、抵押咨询类、还款方式咨询类、审核时长咨询类及还款计划咨询类均为用户意图类型的小类别。例如,意图识别训练模块150将上述用户2的语句处理后输入训练好的意图识别分类模型,识别出用户2的意图是审核时长咨询类,需要了解贷款审核时长。
步骤S40,根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户2。例如,根据上述语句“我提交的借款一星期了,还要多久能审核下来。”业务咨询模块170查询自领域知识库31,得到贷款审核的时长为两个星期,则对话生成模块190结合用户2“愤怒”和“焦虑”的情绪向用户2回复:“尊敬的用户,您好!您别着急,您的贷款一个星期后将审核结束,谢谢!”
在另一个实施例中,若用户2输入的语句为:“我想贷20000元。”,分词及词性标注后的结果为:“我/PN想/VV贷/V20000/M元/RN。/PU”,命名实体识别出该语句中的货币命名实体“20000元”。语句类型识别模块120结合分词、词性判别的结果以及依存句法分析识别出用户2输入的语句的句子类型。例如,根据该用户2输入语句中的“贷”等词语的分析,识别出该语句为陈述句。情感分析模块130结合句子类型,将处理后的语句输入基于词典的情感分析方式和预先训练好的基于深度学习的情感分类模型,识别出用户2的情绪类别。例如,结合句子类型,通过“贷”等词语识别出用户情绪极性为负性,同时分析出此时用户2的主要情绪为“焦虑”。对话引擎分析模块140识别用户2当前处于对话流程中,调用模板引擎模块150查找与当前用户意图相匹配的对话流程模板为贷款申请流程模板,引导用户完成资料填写、资料校验,如果用户具有贷款资格,告知用户贷款金额发放时间、还款日期及还款金额。如果用户不具备贷款资格,提示用户完善资料。
上述实施例提出的智能交互方法,通过对用户输入的语句进行处理,分析语句类型、用户的情绪,之后判断用户是否处于流程对话中,用户处于流程中,则确定对话流程模板,引导用户贷款的进行,若用户不处于流程中,则根据用户不同的意图执行贷款咨询类及闲聊类等相关操作,结合用户情绪为用户生成不同的回复,与用户进行互动,提高自动化水平及用户体验,提升语义理解准确性。
此外,本申请实施例还提出一种计算机可读存储介质,所述计算机可读存储介质中包括智能交互程序10,所述智能交互程序10被处理器执行时实现如下操作:
构建步骤:构建不同业务流程的对话流程模板,并训练意图识别分类模型;
处理步骤:接收用户输入的语句,对语句执行一系列处理,包括预处理、识别语句类型以及情感分析,得到用户的情绪类别;
识别步骤:利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的语句中识别用户意图;
反馈步骤:根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户。
优选地,所述意图识别分类模型为CNN模型,所述意图识别分类模型的训练步骤如下:
整理步骤:整理训练数据,包括获取用户语料以及标注用户语料的意图分类;
训练步骤:利用skip-gram或者CBOW模型训练用户语料,得到低维度的字向量;
转换步骤:通过查找字向量表的方式,将低维度的字向量转换成相应的向量;
抽取步骤:在特征矩阵的维度上,设定窗口的大小,通过卷积、池化操作将所述向量表示式转换成特征向量,对所述特征向量使用Hard Tanh函数进行非线性特征的抽取,经过预设次数的特征抽取之后,获取用户语料的最终特征;
大类训练步骤:利用梯度下降算法迭代,对意图分类中的大类进行有监督的训练;
小类训练步骤:对意图分类中大类的每个小类,使用上述转换步骤、抽取步骤及大类训练步骤的训练方式进行训练。
优选地,所述识别步骤包括:识别用户当前是否在流程对话中,若用户处于流程对话中,则解析对话内容以确定对应的对话流程模板,根据预先构 建的对话流程模板引导对话流程,若用户处于流程对话外,则调用预先训练好的意图识别分类模型识别用户意图。
优选地,所述识别步骤还包括:若用户之前已触发流程对话,并且在预设时间内再次触发对话,则判断用户处于流程对话中,若用户第一次触发对话或者超过预设时间后再次触发对话,则识别用户当前是否在流程对话中。
优选地,所述识别步骤还包括:
将所述语句输入意图识别分类模型,得到该语句的意图识别分数;
当意图识别分数高于或等于第一阈值时,判断用户意图分类为业务咨询类,当意图识别分数低于第一阈值且高于第二阈值时,判断用户意图分类为闲聊类中的领域相关问题类,当意图识别分数低于或等于第二阈值时,判断用户意图分类为闲聊类中的其他问题类。
优选地,所述预处理包括:分词、词性标注、NER、指代消歧,以及相似词语扩展。
本申请之计算机可读存储介质的具体实施方式与上述智能交互方法的具体实施方式大致相同,在此不再赘述。
上述本申请实施例序号仅仅为了描述,不代表实施例的优劣。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来,该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中,包括若干指令用以使得一台终端设备(可以是手机,计算机,服务器,或者网络设备等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例,并非因此限制本申请的专利范围,凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换,或直接或间接运用在其他相关的技术领域,均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种智能交互方法,应用于电子装置,其特征在于,所述方法包括:
    构建步骤:构建不同业务流程的对话流程模板,并训练意图识别分类模型;
    处理步骤:接收用户输入的语句,对语句执行一系列处理,包括预处理、识别语句类型以及情感分析,得到用户的情绪类别;
    识别步骤:利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的语句中识别用户意图;
    反馈步骤:根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户。
  2. 根据权利要求1所述的智能交互方法,其特征在于,所述意图识别分类模型为卷积神经网络模型,所述意图识别分类模型的训练步骤如下:
    整理步骤:整理训练数据,包括获取用户语料以及标注用户语料的意图分类;
    训练步骤:利用skip-gram或者连续词袋模型训练用户语料,得到低维度的字向量;
    转换步骤:通过查找字向量表的方式,将低维度的字向量转换成相应的向量;
    抽取步骤:在特征矩阵的维度上,设定窗口的大小,通过卷积、池化操作将所述向量表示式转换成特征向量,对所述特征向量使用Hard Tanh函数进行非线性特征的抽取,经过预设次数的特征抽取之后,获取用户语料的最终特征;
    大类训练步骤:利用梯度下降算法迭代,对意图分类中的大类进行有监督的训练;
    小类训练步骤:对意图分类中大类的每个小类,使用上述转换步骤、抽取步骤及大类训练步骤的训练方式进行训练。
  3. 根据权利要求1所述的智能交互方法,其特征在于,所述识别步骤包括:识别用户当前是否在流程对话中,若用户处于流程对话中,则解析对话内容以确定对应的对话流程模板,根据预先构建的对话流程模板引导对话流程,若用户处于流程对话外,则调用预先训练好的意图识别分类模型识别用 户意图。
  4. 根据权利要求1或3所述的智能交互方法,其特征在于,所述识别步骤还包括:若用户之前已触发流程对话,并且在预设时间内再次触发对话,则判断用户处于流程对话中,若用户第一次触发对话或者超过预设时间后再次触发对话,则识别用户当前是否在流程对话中。
  5. 根据权利要求3所述的智能交互方法,其特征在于,所述识别步骤还包括:
    将所述语句输入意图识别分类模型,得到该语句的意图识别分数;
    当意图识别分数高于或等于第一阈值时,判断用户意图分类为业务咨询类,当意图识别分数低于第一阈值且高于第二阈值时,判断用户意图分类为闲聊类中的领域相关问题类,当意图识别分数低于或等于第二阈值时,判断用户意图分类为闲聊类中的其他问题类。
  6. 根据权利要求1所述的智能交互方法,其特征在于,所述预处理包括:分词、词性标注、命名实体识别、指代消歧,以及相似词语扩展。
  7. 根据权利要求6所述的智能交互方法,其特征在于,所述相似词语扩展是指利用Word2vec技术进行相似词语的扩展。
  8. 一种电子装置,其特征在于,所述电子装置包括:存储器、处理器及显示器,所述存储器上存储有智能交互程序,所述智能交互程序被所述处理器执行,可实现如下步骤:
    构建步骤:构建不同业务流程的对话流程模板,并训练意图识别分类模型;
    处理步骤:接收用户输入的语句,对语句执行一系列处理,包括预处理、识别语句类型以及情感分析,得到用户的情绪类别;
    识别步骤:利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的语句中识别用户意图;
    反馈步骤:根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户。
  9. 根据权利要求8所述的电子装置,其特征在于,所述意图识别分类模型为卷积神经网络模型,所述意图识别分类模型的训练步骤如下:
    整理步骤:整理训练数据,包括获取用户语料以及标注用户语料的意图 分类;
    训练步骤:利用skip-gram或者连续词袋模型训练用户语料,得到低维度的字向量;
    转换步骤:通过查找字向量表的方式,将低维度的字向量转换成相应的向量;
    抽取步骤:在特征矩阵的维度上,设定窗口的大小,通过卷积、池化操作将所述向量表示式转换成特征向量,对所述特征向量使用Hard Tanh函数进行非线性特征的抽取,经过预设次数的特征抽取之后,获取用户语料的最终特征;
    大类训练步骤:利用梯度下降算法迭代,对意图分类中的大类进行有监督的训练;
    小类训练步骤:对意图分类中大类的每个小类,使用上述转换步骤、抽取步骤及大类训练步骤的训练方式进行训练。
  10. 根据权利要求8所述的电子装置,其特征在于,所述识别步骤包括:识别用户当前是否在流程对话中,若用户处于流程对话中,则解析对话内容以确定对应的对话流程模板,根据预先构建的对话流程模板引导对话流程,若用户处于流程对话外,则调用预先训练好的意图识别分类模型识别用户意图。
  11. 根据权利要求8或10所述的电子装置,其特征在于,所述识别步骤还包括:若用户之前已触发流程对话,并且在预设时间内再次触发对话,则判断用户处于流程对话中,若用户第一次触发对话或者超过预设时间后再次触发对话,则识别用户当前是否在流程对话中。
  12. 根据权利要求10所述的电子装置,其特征在于,所述识别步骤还包括:
    将所述语句输入意图识别分类模型,得到该语句的意图识别分数;
    当意图识别分数高于或等于第一阈值时,判断用户意图分类为业务咨询类,当意图识别分数低于第一阈值且高于第二阈值时,判断用户意图分类为闲聊类中的领域相关问题类,当意图识别分数低于或等于第二阈值时,判断用户意图分类为闲聊类中的其他问题类。
  13. 根据权利要求8所述的电子装置,其特征在于,所述预处理包括: 分词、词性标注、命名实体识别、指代消歧,以及相似词语扩展。
  14. 根据权利要求13所述的电子装置,其特征在于,所述相似词语扩展是指利用Word2vec技术进行相似词语的扩展。
  15. 一种计算机可读存储介质,其特征在于,所述计算机可读存储介质中包括智能交互程序,所述统智能交互程序被处理器执行时实现如下步骤:
    构建步骤:构建不同业务流程的对话流程模板,并训练意图识别分类模型;
    处理步骤:接收用户输入的语句,对语句执行一系列处理,包括预处理、识别语句类型以及情感分析,得到用户的情绪类别;
    识别步骤:利用预先构建的对话流程模板及训练好的意图识别分类模型,从处理后的语句中识别用户意图;
    反馈步骤:根据识别到的用户意图查询相应的知识库,将查询结果结合用户情绪类别生成回复,反馈给用户。
  16. 根据权利要求15所述的计算机可读存储介质,其特征在于,所述意图识别分类模型为卷积神经网络模型,所述意图识别分类模型的训练步骤如下:
    整理步骤:整理训练数据,包括获取用户语料以及标注用户语料的意图分类;
    训练步骤:利用skip-gram或者连续词袋模型训练用户语料,得到低维度的字向量;
    转换步骤:通过查找字向量表的方式,将低维度的字向量转换成相应的向量;
    抽取步骤:在特征矩阵的维度上,设定窗口的大小,通过卷积、池化操作将所述向量表示式转换成特征向量,对所述特征向量使用Hard Tanh函数进行非线性特征的抽取,经过预设次数的特征抽取之后,获取用户语料的最终特征;
    大类训练步骤:利用梯度下降算法迭代,对意图分类中的大类进行有监督的训练;
    小类训练步骤:对意图分类中大类的每个小类,使用上述转换步骤、抽取步骤及大类训练步骤的训练方式进行训练。
  17. 根据权利要求15所述的计算机可读存储介质,其特征在于,所述识别步骤包括:识别用户当前是否在流程对话中,若用户处于流程对话中,则解析对话内容以确定对应的对话流程模板,根据预先构建的对话流程模板引导对话流程,若用户处于流程对话外,则调用预先训练好的意图识别分类模型识别用户意图。
  18. 根据权利要求15或17所述的计算机可读存储介质,其特征在于,所述识别步骤还包括:若用户之前已触发流程对话,并且在预设时间内再次触发对话,则判断用户处于流程对话中,若用户第一次触发对话或者超过预设时间后再次触发对话,则识别用户当前是否在流程对话中。
  19. 根据权利要求17所述的计算机可读存储介质,其特征在于,所述识别步骤还包括:
    将所述语句输入意图识别分类模型,得到该语句的意图识别分数;
    当意图识别分数高于或等于第一阈值时,判断用户意图分类为业务咨询类,当意图识别分数低于第一阈值且高于第二阈值时,判断用户意图分类为闲聊类中的领域相关问题类,当意图识别分数低于或等于第二阈值时,判断用户意图分类为闲聊类中的其他问题类。
  20. 根据权利要求15所述的计算机可读存储介质,其特征在于,所述预处理包括:分词、词性标注、命名实体识别、指代消歧,以及相似词语扩展。
PCT/CN2018/084407 2018-02-09 2018-04-25 智能交互方法、电子装置及存储介质 WO2019153522A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810134194.2A CN108427722A (zh) 2018-02-09 2018-02-09 智能交互方法、电子装置及存储介质
CN201810134194.2 2018-02-09

Publications (1)

Publication Number Publication Date
WO2019153522A1 true WO2019153522A1 (zh) 2019-08-15

Family

ID=63156692

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/084407 WO2019153522A1 (zh) 2018-02-09 2018-04-25 智能交互方法、电子装置及存储介质

Country Status (2)

Country Link
CN (1) CN108427722A (zh)
WO (1) WO2019153522A1 (zh)

Cited By (49)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866099A (zh) * 2019-10-30 2020-03-06 南昌众荟智盈信息技术有限公司 一种基于智能音箱语音交互的智能管家服务方法及系统
CN111274390A (zh) * 2020-01-15 2020-06-12 深圳前海微众银行股份有限公司 一种基于对话数据的情感原因确定方法及装置
CN111309859A (zh) * 2020-01-21 2020-06-19 上饶市中科院云计算中心大数据研究院 一种景区网络口碑情感分析方法及装置
CN111324712A (zh) * 2020-02-18 2020-06-23 山东汇贸电子口岸有限公司 对话回复方法及服务端
CN111368538A (zh) * 2020-02-29 2020-07-03 平安科技(深圳)有限公司 语音交互方法、系统、终端及计算机可读存储介质
CN111460124A (zh) * 2020-04-21 2020-07-28 文思海辉智科科技有限公司 智能交互方法、装置及机器人
CN111666389A (zh) * 2020-04-21 2020-09-15 文思海辉智科科技有限公司 对话数据处理方法、装置、计算机设备和存储介质
CN111753552A (zh) * 2020-07-02 2020-10-09 浙江百应科技有限公司 一种基于nlp的训练模式与识别模式动态切换的方法
CN111858890A (zh) * 2020-07-23 2020-10-30 平安证券股份有限公司 基于机器人的咨询方法、咨询机器人、计算机设备及介质
CN111949793A (zh) * 2020-08-13 2020-11-17 深圳市欢太科技有限公司 用户意图识别方法、装置及终端设备
CN112015919A (zh) * 2020-09-15 2020-12-01 重庆广播电视大学重庆工商职业学院 一种基于学习辅助知识图谱的对话管理方法
CN112036154A (zh) * 2020-08-31 2020-12-04 康键信息技术(深圳)有限公司 基于问诊对话的电子病历生成方法、装置和计算机设备
CN112071436A (zh) * 2020-09-24 2020-12-11 上海依图网络科技有限公司 针对疫情的智能交互方法、装置及存储介质
CN112131359A (zh) * 2020-09-04 2020-12-25 交通银行股份有限公司太平洋信用卡中心 一种基于图形化编排智能策略的意图识别方法及电子设备
CN112164394A (zh) * 2020-09-10 2021-01-01 北京三快在线科技有限公司 一种信息交互方法、装置、存储介质及电子设备
CN112182175A (zh) * 2020-09-25 2021-01-05 北京字节跳动网络技术有限公司 智能问答方法、装置、设备及可读存储介质
CN112270189A (zh) * 2020-11-12 2021-01-26 佰聆数据股份有限公司 一种提问式的分析节点生成方法、系统及存储介质
CN112597288A (zh) * 2020-12-23 2021-04-02 北京百度网讯科技有限公司 人机交互方法、装置、设备及存储介质
CN112612878A (zh) * 2020-12-17 2021-04-06 大唐融合通信股份有限公司 一种客服信息提供方法、电子设备及装置
CN112650834A (zh) * 2020-12-25 2021-04-13 竹间智能科技(上海)有限公司 一种意图模型训练方法及装置
CN112667792A (zh) * 2020-12-29 2021-04-16 招联消费金融有限公司 人机对话数据处理方法、装置、计算机设备和存储介质
CN112686674A (zh) * 2020-12-25 2021-04-20 科讯嘉联信息技术有限公司 一种客服对话工单总结方法
CN112905893A (zh) * 2021-03-22 2021-06-04 北京百度网讯科技有限公司 搜索意图识别模型的训练方法、搜索意图识别方法及装置
CN113033160A (zh) * 2019-12-09 2021-06-25 阿里巴巴集团控股有限公司 对话的意图分类方法及设备和生成意图分类模型的方法
CN113127621A (zh) * 2021-04-28 2021-07-16 平安国际智慧城市科技股份有限公司 对话模块的推送方法、装置、设备及存储介质
CN113158680A (zh) * 2021-03-23 2021-07-23 北京新方通信技术有限公司 一种语料处理及意图识别的方法和装置
CN113159901A (zh) * 2021-04-29 2021-07-23 天津狮拓信息技术有限公司 融资租赁业务会话的实现方法和装置
CN113157873A (zh) * 2021-01-25 2021-07-23 北京海致星图科技有限公司 一种基于模板匹配和深度学习的知识库问答系统构建方法
CN113360639A (zh) * 2020-03-06 2021-09-07 上海卓繁信息技术股份有限公司 一种短文本情绪分类方法,装置及存储设备
CN113420542A (zh) * 2021-06-23 2021-09-21 平安科技(深圳)有限公司 对话生成方法、装置、电子设备及存储介质
CN113420115A (zh) * 2021-06-22 2021-09-21 平安科技(深圳)有限公司 基于人机语音对话的信息处理方法及装置
CN113449086A (zh) * 2020-03-24 2021-09-28 北京有限元科技有限公司 对话管理方法、装置以及存储介质
CN113449113A (zh) * 2020-03-27 2021-09-28 京东数字科技控股有限公司 一种知识图谱构建方法、装置、电子设备及存储介质
CN113569022A (zh) * 2021-06-30 2021-10-29 四川新网银行股份有限公司 一种基于级联式搜索的对话机器人应答引擎实现方法
CN113569021A (zh) * 2021-06-29 2021-10-29 杭州摸象大数据科技有限公司 用户分类的方法、计算机设备和可读存储介质
CN113609280A (zh) * 2021-08-05 2021-11-05 云南大学 基于元学习的多领域对话生成方法、装置、设备及介质
CN113672715A (zh) * 2021-08-20 2021-11-19 上海大参林医疗健康科技有限公司 一种意图识别系统及方法
CN113707303A (zh) * 2021-08-30 2021-11-26 康键信息技术(深圳)有限公司 基于知识图谱的医疗问题解答方法、装置、设备及介质
CN113761139A (zh) * 2020-08-10 2021-12-07 北京沃东天骏信息技术有限公司 一种意图预判的方法和装置
CN113780012A (zh) * 2021-09-30 2021-12-10 东南大学 基于预训练语言模型的抑郁症访谈对话生成方法
CN113836278A (zh) * 2021-08-13 2021-12-24 北京百度网讯科技有限公司 通用对话模型的训练与对话生成方法、装置
CN113837766A (zh) * 2021-10-08 2021-12-24 支付宝(杭州)信息技术有限公司 风险识别方法、装置和电子设备
CN114880472A (zh) * 2022-04-28 2022-08-09 支付宝(杭州)信息技术有限公司 数据处理方法、装置及设备
CN115033676A (zh) * 2022-06-22 2022-09-09 支付宝(杭州)信息技术有限公司 一种意图识别模型训练及用户意图识别方法及装置
CN115840808A (zh) * 2022-12-27 2023-03-24 广州汉申科技中介服务有限公司 科技项目咨询方法、装置、服务器及计算机可读存储介质
CN116307663A (zh) * 2023-02-16 2023-06-23 广州力麒智能科技有限公司 一种智能自助政务服务系统、政务服务通办的方法及终端
CN116303982A (zh) * 2023-05-24 2023-06-23 杭州实在智能科技有限公司 基于rpa和自学习机制的智能应答与业务处理方法及系统
CN116756277A (zh) * 2023-04-20 2023-09-15 海尔优家智能科技(北京)有限公司 基于目标生成式预训练gpt模型的交互语句的处理方法
CN117252211A (zh) * 2023-10-18 2023-12-19 中建八局第一数字科技有限公司 提问意图识别模型的构建、提问意图识别方法及装置

Families Citing this family (69)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109461039A (zh) * 2018-08-28 2019-03-12 厦门快商通信息技术有限公司 一种文本处理方法及智能客服方法
CN109145104B (zh) * 2018-09-29 2023-04-18 北京百度网讯科技有限公司 用于对话交互的方法和装置
CN109522927A (zh) * 2018-10-09 2019-03-26 北京奔影网络科技有限公司 用于用户消息的情感分析方法及装置
CN109408821B (zh) * 2018-10-22 2020-09-04 腾讯科技(深圳)有限公司 一种语料生成方法、装置、计算设备及存储介质
CN109614464B (zh) * 2018-10-31 2023-10-27 创新先进技术有限公司 用于业务问题识别的方法及装置
CN109492104B (zh) * 2018-11-09 2023-01-31 北京汇钧科技有限公司 意图分类模型的训练方法、分类方法、系统、设备和介质
CN111178081B (zh) * 2018-11-09 2023-07-21 中移(杭州)信息技术有限公司 语义识别的方法、服务器、电子设备及计算机存储介质
CN109543182B (zh) * 2018-11-15 2023-04-07 广东电网有限责任公司信息中心 一种基于solr引擎的电力企业多轮交互语义分析方法
CN109522556B (zh) * 2018-11-16 2024-03-12 北京九狐时代智能科技有限公司 一种意图识别方法及装置
CN111274382A (zh) * 2018-11-20 2020-06-12 北京京东尚科信息技术有限公司 文本分类方法、装置、设备及存储介质
CN111210824B (zh) * 2018-11-21 2023-04-07 深圳绿米联创科技有限公司 语音信息处理方法、装置、电子设备及存储介质
CN109829038A (zh) * 2018-12-11 2019-05-31 平安科技(深圳)有限公司 基于深度学习的问答反馈方法、装置、设备及存储介质
CN109753565A (zh) * 2018-12-27 2019-05-14 厦门智融合科技有限公司 知识产权智能服务方法及系统
CN109859747B (zh) * 2018-12-29 2021-05-04 北京百度网讯科技有限公司 语音交互方法、设备以及存储介质
CN109935242A (zh) * 2019-01-10 2019-06-25 上海言通网络科技有限公司 可打断式语音处理系统和方法
CN109753561B (zh) * 2019-01-16 2021-04-27 长安汽车金融有限公司 一种自动回复的生成方法及装置
CN111563209B (zh) * 2019-01-29 2023-06-30 株式会社理光 一种意图识别的方法、装置及计算机可读存储介质
CN109961786B (zh) * 2019-01-31 2023-04-14 平安科技(深圳)有限公司 基于语音分析的产品推荐方法、装置、设备和存储介质
CN111767735A (zh) * 2019-03-26 2020-10-13 北京京东尚科信息技术有限公司 执行任务的方法、装置及计算机可读存储介质
CN111782775B (zh) * 2019-04-04 2023-09-01 百度在线网络技术(北京)有限公司 对话方法、装置、设备和介质
CN110008331B (zh) * 2019-04-15 2021-09-14 腾讯科技(深圳)有限公司 信息展示方法、装置、电子设备及计算机可读存储介质
CN110110058A (zh) * 2019-04-22 2019-08-09 南京硅基智能科技有限公司 一种外呼机器人的意向打分方法及系统
CN110232114A (zh) * 2019-05-06 2019-09-13 平安科技(深圳)有限公司 语句意图识别方法、装置及计算机可读存储介质
CN110377708B (zh) * 2019-06-03 2021-10-08 广东幽澜机器人科技有限公司 一种多情景对话切换方法及装置
CN110363074B (zh) * 2019-06-03 2021-03-30 华南理工大学 一种针对复杂抽象化事物的类人化识别交互方法
CN110188714A (zh) * 2019-06-04 2019-08-30 言图科技有限公司 一种在聊天场景下实现财务管理的方法、系统及存储介质
CN110379445A (zh) * 2019-06-20 2019-10-25 深圳壹账通智能科技有限公司 基于情绪分析的业务处理方法、装置、设备及存储介质
CN110276403B (zh) * 2019-06-25 2022-02-18 北京百度网讯科技有限公司 模型建立方法和装置
CN110263346B (zh) * 2019-06-27 2023-01-24 卓尔智联(武汉)研究院有限公司 基于小样本学习的语意分析方法、电子设备及存储介质
CN110442701B (zh) * 2019-08-15 2022-08-05 思必驰科技股份有限公司 语音对话处理方法及装置
CN112612876A (zh) * 2019-09-19 2021-04-06 北京国双科技有限公司 用户咨询的应对处理方法、装置、存储介质及电子设备
CN110704623A (zh) * 2019-09-27 2020-01-17 安徽咪鼠科技有限公司 基于Rasa_Nlu框架提高实体识别率的方法、装置、系统和存储介质
CN110688459A (zh) * 2019-09-29 2020-01-14 联想(北京)有限公司 一种智能交互方法及智能设备
CN110647617B (zh) * 2019-09-29 2022-04-19 百度在线网络技术(北京)有限公司 对话引导模型的训练样本构建方法和生成模型的方法
CN110931002B (zh) * 2019-10-12 2023-06-20 平安科技(深圳)有限公司 人机交互方法、装置、计算机设备和存储介质
CN110990543A (zh) * 2019-10-18 2020-04-10 平安科技(深圳)有限公司 智能对话的生成方法、装置、计算机设备及计算机存储介质
TWI752367B (zh) * 2019-10-30 2022-01-11 中國信託商業銀行股份有限公司 基於自然語言處理的智能對話管理方法及系統
CN110826339B (zh) * 2019-10-31 2024-03-01 联想(北京)有限公司 行为识别方法、装置、电子设备以及介质
CN111026843B (zh) * 2019-12-02 2023-03-14 北京智乐瑟维科技有限公司 一种人工智能语音外呼方法、系统及存储介质
CN111159346A (zh) * 2019-12-27 2020-05-15 深圳物控智联科技有限公司 基于意图识别的智能答疑方法、服务器及存储介质
CN111221950A (zh) * 2019-12-30 2020-06-02 航天信息股份有限公司 一种用户弱感情的分析方法及装置
CN111177358B (zh) * 2019-12-31 2023-05-12 华为技术有限公司 意图识别方法、服务器及存储介质
CN111309306B (zh) * 2020-02-24 2023-07-28 福建天晴数码有限公司 人机交互的对话管理系统
CN111460123B (zh) * 2020-04-07 2020-10-20 中国搜索信息科技股份有限公司 一种针对青少年聊天机器人的对话意图识别方法及装置
CN111581623B (zh) * 2020-05-09 2023-12-19 深圳物控智联科技有限公司 智能数据交互方法、装置、电子设备及存储介质
CN111737412A (zh) * 2020-05-21 2020-10-02 电子科技大学 基于自然语言处理和知识图谱的公民来访引导方法
CN115803748A (zh) * 2020-05-28 2023-03-14 深圳市欢太数字科技有限公司 文本模板提取方法、电子设备和存储介质
CN113779962B (zh) * 2020-06-10 2024-02-02 阿里巴巴集团控股有限公司 数据处理方法、装置、设备和存储介质
CN111782792A (zh) * 2020-08-05 2020-10-16 支付宝(杭州)信息技术有限公司 用于信息处理的方法和装置
CN112131357A (zh) * 2020-08-21 2020-12-25 国网浙江省电力有限公司杭州供电公司 一种基于智能对话模型的用户意图识别方法及装置
CN112035640A (zh) * 2020-08-31 2020-12-04 重庆长安汽车股份有限公司 一种基于智能问答机器人的精细化问答方法、存储介质和智能设备
CN112035620B (zh) * 2020-08-31 2023-02-14 康键信息技术(深圳)有限公司 医疗查询系统的问答管理方法、装置、设备及存储介质
CN114255044A (zh) * 2020-09-11 2022-03-29 四川大学 基于跨媒体分析的智能客服技术
CN112380240A (zh) * 2020-11-17 2021-02-19 深圳壹账通智能科技有限公司 基于语义识别的数据查询方法、装置、设备及存储介质
CN112417121A (zh) * 2020-11-20 2021-02-26 平安普惠企业管理有限公司 客户意图识别方法、装置、计算机设备及存储介质
CN112487166A (zh) * 2020-12-04 2021-03-12 浪潮云信息技术股份公司 一种事项智能问答的实现方法
CN112597287B (zh) * 2020-12-15 2024-03-15 深圳市优必选科技股份有限公司 一种语句处理方法、语句处理装置及智能设备
CN112464674A (zh) * 2020-12-16 2021-03-09 四川长虹电器股份有限公司 一种字级别的文本意图识别方法
CN112765313B (zh) * 2020-12-31 2022-06-28 太原理工大学 一种基于原文和评论信息分析算法的虚假信息检测方法
CN112765331B (zh) * 2020-12-31 2022-11-18 杭州摸象大数据科技有限公司 对话知识模板构建方法、装置、计算机设备及存储介质
CN113792124A (zh) * 2021-01-15 2021-12-14 北京沃东天骏信息技术有限公司 智能应答方法、装置、电子设备及存储介质
CN113010784B (zh) * 2021-03-17 2024-02-06 北京十一贝科技有限公司 用于生成预测信息的方法、装置、电子设备和介质
CN112951233A (zh) * 2021-03-30 2021-06-11 平安科技(深圳)有限公司 语音问答方法、装置、电子设备及可读存储介质
CN113157895B (zh) * 2021-05-25 2023-07-25 中国平安人寿保险股份有限公司 智能对话分析方法、系统、计算机设备和存储介质
CN113282755A (zh) * 2021-06-11 2021-08-20 上海寻梦信息技术有限公司 对话型文本分类方法、系统、设备及存储介质
CN113626568A (zh) * 2021-07-30 2021-11-09 平安普惠企业管理有限公司 机器人的人机对话控制方法、装置、计算机设备及介质
CN113779229A (zh) * 2021-08-31 2021-12-10 康键信息技术(深圳)有限公司 用户需求识别匹配方法、装置、设备和可读存储介质
CN113743126B (zh) * 2021-11-08 2022-06-14 北京博瑞彤芸科技股份有限公司 一种基于用户情绪的智能交互方法和装置
CN117078270B (zh) * 2023-10-17 2024-02-02 彩讯科技股份有限公司 用于网络产品营销的智能交互方法和装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106296243A (zh) * 2015-05-22 2017-01-04 阿里巴巴集团控股有限公司 业务实现方法及装置
CN106326984A (zh) * 2016-08-09 2017-01-11 北京京东尚科信息技术有限公司 用户意图识别方法和装置、自动应答系统
CN106649694A (zh) * 2016-12-19 2017-05-10 北京云知声信息技术有限公司 语音交互中确定用户意图的方法及装置
CN107562816A (zh) * 2017-08-16 2018-01-09 深圳狗尾草智能科技有限公司 用户意图自动识别方法及装置

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104765791A (zh) * 2015-03-24 2015-07-08 北京搜狗科技发展有限公司 一种信息输入方法和装置
CN106897471A (zh) * 2015-12-21 2017-06-27 北京航空航天大学 设计意图树模板化方法
CN106484916A (zh) * 2016-11-04 2017-03-08 上海智臻智能网络科技股份有限公司 一种招生问答方法及装置
CN106773923B (zh) * 2016-11-30 2020-04-21 北京光年无限科技有限公司 面向机器人的多模态情感数据交互方法及装置
CN106909220A (zh) * 2017-02-21 2017-06-30 山东师范大学 一种适用于触控的视线交互方法
CN107135143A (zh) * 2017-03-27 2017-09-05 厦门快商通科技股份有限公司 多聊天机器人切换系统及其对话方法
CN107657852B (zh) * 2017-11-14 2023-09-22 翟奕雲 基于人脸识别的幼儿教学机器人、教学系统、存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106296243A (zh) * 2015-05-22 2017-01-04 阿里巴巴集团控股有限公司 业务实现方法及装置
CN106326984A (zh) * 2016-08-09 2017-01-11 北京京东尚科信息技术有限公司 用户意图识别方法和装置、自动应答系统
CN106649694A (zh) * 2016-12-19 2017-05-10 北京云知声信息技术有限公司 语音交互中确定用户意图的方法及装置
CN107562816A (zh) * 2017-08-16 2018-01-09 深圳狗尾草智能科技有限公司 用户意图自动识别方法及装置

Cited By (76)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866099A (zh) * 2019-10-30 2020-03-06 南昌众荟智盈信息技术有限公司 一种基于智能音箱语音交互的智能管家服务方法及系统
CN110866099B (zh) * 2019-10-30 2023-05-09 上海益商网络科技有限公司 一种基于智能音箱语音交互的智能管家服务方法及系统
CN113033160A (zh) * 2019-12-09 2021-06-25 阿里巴巴集团控股有限公司 对话的意图分类方法及设备和生成意图分类模型的方法
CN111274390A (zh) * 2020-01-15 2020-06-12 深圳前海微众银行股份有限公司 一种基于对话数据的情感原因确定方法及装置
CN111274390B (zh) * 2020-01-15 2023-10-27 深圳前海微众银行股份有限公司 一种基于对话数据的情感原因确定方法及装置
CN111309859A (zh) * 2020-01-21 2020-06-19 上饶市中科院云计算中心大数据研究院 一种景区网络口碑情感分析方法及装置
CN111309859B (zh) * 2020-01-21 2023-07-07 上饶市中科院云计算中心大数据研究院 一种景区网络口碑情感分析方法及装置
CN111324712A (zh) * 2020-02-18 2020-06-23 山东汇贸电子口岸有限公司 对话回复方法及服务端
CN111368538A (zh) * 2020-02-29 2020-07-03 平安科技(深圳)有限公司 语音交互方法、系统、终端及计算机可读存储介质
CN111368538B (zh) * 2020-02-29 2023-10-24 平安科技(深圳)有限公司 语音交互方法、系统、终端及计算机可读存储介质
CN113360639A (zh) * 2020-03-06 2021-09-07 上海卓繁信息技术股份有限公司 一种短文本情绪分类方法,装置及存储设备
CN113449086A (zh) * 2020-03-24 2021-09-28 北京有限元科技有限公司 对话管理方法、装置以及存储介质
CN113449113A (zh) * 2020-03-27 2021-09-28 京东数字科技控股有限公司 一种知识图谱构建方法、装置、电子设备及存储介质
CN111666389A (zh) * 2020-04-21 2020-09-15 文思海辉智科科技有限公司 对话数据处理方法、装置、计算机设备和存储介质
CN111460124A (zh) * 2020-04-21 2020-07-28 文思海辉智科科技有限公司 智能交互方法、装置及机器人
CN111460124B (zh) * 2020-04-21 2023-08-25 文思海辉智科科技有限公司 智能交互方法、装置及机器人
CN111753552A (zh) * 2020-07-02 2020-10-09 浙江百应科技有限公司 一种基于nlp的训练模式与识别模式动态切换的方法
CN111753552B (zh) * 2020-07-02 2024-04-19 浙江百应科技有限公司 一种基于nlp的训练模式与识别模式动态切换的方法
CN111858890B (zh) * 2020-07-23 2023-07-21 平安证券股份有限公司 基于机器人的咨询方法、咨询机器人、计算机设备及介质
CN111858890A (zh) * 2020-07-23 2020-10-30 平安证券股份有限公司 基于机器人的咨询方法、咨询机器人、计算机设备及介质
CN113761139A (zh) * 2020-08-10 2021-12-07 北京沃东天骏信息技术有限公司 一种意图预判的方法和装置
CN111949793B (zh) * 2020-08-13 2024-02-27 深圳市欢太科技有限公司 用户意图识别方法、装置及终端设备
CN111949793A (zh) * 2020-08-13 2020-11-17 深圳市欢太科技有限公司 用户意图识别方法、装置及终端设备
CN112036154B (zh) * 2020-08-31 2023-05-23 康键信息技术(深圳)有限公司 基于问诊对话的电子病历生成方法、装置和计算机设备
CN112036154A (zh) * 2020-08-31 2020-12-04 康键信息技术(深圳)有限公司 基于问诊对话的电子病历生成方法、装置和计算机设备
CN112131359A (zh) * 2020-09-04 2020-12-25 交通银行股份有限公司太平洋信用卡中心 一种基于图形化编排智能策略的意图识别方法及电子设备
CN112164394A (zh) * 2020-09-10 2021-01-01 北京三快在线科技有限公司 一种信息交互方法、装置、存储介质及电子设备
CN112015919A (zh) * 2020-09-15 2020-12-01 重庆广播电视大学重庆工商职业学院 一种基于学习辅助知识图谱的对话管理方法
CN112071436A (zh) * 2020-09-24 2020-12-11 上海依图网络科技有限公司 针对疫情的智能交互方法、装置及存储介质
CN112182175A (zh) * 2020-09-25 2021-01-05 北京字节跳动网络技术有限公司 智能问答方法、装置、设备及可读存储介质
CN112270189A (zh) * 2020-11-12 2021-01-26 佰聆数据股份有限公司 一种提问式的分析节点生成方法、系统及存储介质
CN112270189B (zh) * 2020-11-12 2023-07-18 佰聆数据股份有限公司 一种提问式的分析节点生成方法、系统及存储介质
CN112612878A (zh) * 2020-12-17 2021-04-06 大唐融合通信股份有限公司 一种客服信息提供方法、电子设备及装置
CN112597288B (zh) * 2020-12-23 2023-07-25 北京百度网讯科技有限公司 人机交互方法、装置、设备及存储介质
CN112597288A (zh) * 2020-12-23 2021-04-02 北京百度网讯科技有限公司 人机交互方法、装置、设备及存储介质
CN112650834B (zh) * 2020-12-25 2023-10-03 竹间智能科技(上海)有限公司 一种意图模型训练方法及装置
CN112650834A (zh) * 2020-12-25 2021-04-13 竹间智能科技(上海)有限公司 一种意图模型训练方法及装置
CN112686674A (zh) * 2020-12-25 2021-04-20 科讯嘉联信息技术有限公司 一种客服对话工单总结方法
CN112667792B (zh) * 2020-12-29 2024-03-22 招联消费金融股份有限公司 人机对话数据处理方法、装置、计算机设备和存储介质
CN112667792A (zh) * 2020-12-29 2021-04-16 招联消费金融有限公司 人机对话数据处理方法、装置、计算机设备和存储介质
CN113157873A (zh) * 2021-01-25 2021-07-23 北京海致星图科技有限公司 一种基于模板匹配和深度学习的知识库问答系统构建方法
CN112905893A (zh) * 2021-03-22 2021-06-04 北京百度网讯科技有限公司 搜索意图识别模型的训练方法、搜索意图识别方法及装置
CN112905893B (zh) * 2021-03-22 2024-01-12 北京百度网讯科技有限公司 搜索意图识别模型的训练方法、搜索意图识别方法及装置
CN113158680B (zh) * 2021-03-23 2024-05-07 北京新方通信技术有限公司 一种语料处理及意图识别的方法和装置
CN113158680A (zh) * 2021-03-23 2021-07-23 北京新方通信技术有限公司 一种语料处理及意图识别的方法和装置
CN113127621A (zh) * 2021-04-28 2021-07-16 平安国际智慧城市科技股份有限公司 对话模块的推送方法、装置、设备及存储介质
CN113159901A (zh) * 2021-04-29 2021-07-23 天津狮拓信息技术有限公司 融资租赁业务会话的实现方法和装置
CN113420115A (zh) * 2021-06-22 2021-09-21 平安科技(深圳)有限公司 基于人机语音对话的信息处理方法及装置
CN113420115B (zh) * 2021-06-22 2024-03-22 平安科技(深圳)有限公司 基于人机语音对话的信息处理方法及装置
CN113420542B (zh) * 2021-06-23 2023-04-18 平安科技(深圳)有限公司 对话生成方法、装置、电子设备及存储介质
CN113420542A (zh) * 2021-06-23 2021-09-21 平安科技(深圳)有限公司 对话生成方法、装置、电子设备及存储介质
CN113569021A (zh) * 2021-06-29 2021-10-29 杭州摸象大数据科技有限公司 用户分类的方法、计算机设备和可读存储介质
CN113569021B (zh) * 2021-06-29 2023-08-04 杭州摸象大数据科技有限公司 用户分类的方法、计算机设备和可读存储介质
CN113569022B (zh) * 2021-06-30 2023-06-09 四川新网银行股份有限公司 一种基于级联式搜索的对话机器人应答引擎实现方法
CN113569022A (zh) * 2021-06-30 2021-10-29 四川新网银行股份有限公司 一种基于级联式搜索的对话机器人应答引擎实现方法
CN113609280A (zh) * 2021-08-05 2021-11-05 云南大学 基于元学习的多领域对话生成方法、装置、设备及介质
CN113609280B (zh) * 2021-08-05 2023-10-13 云南大学 基于元学习的多领域对话生成方法、装置、设备及介质
CN113836278B (zh) * 2021-08-13 2023-08-11 北京百度网讯科技有限公司 通用对话模型的训练与对话生成方法、装置
CN113836278A (zh) * 2021-08-13 2021-12-24 北京百度网讯科技有限公司 通用对话模型的训练与对话生成方法、装置
CN113672715A (zh) * 2021-08-20 2021-11-19 上海大参林医疗健康科技有限公司 一种意图识别系统及方法
CN113707303A (zh) * 2021-08-30 2021-11-26 康键信息技术(深圳)有限公司 基于知识图谱的医疗问题解答方法、装置、设备及介质
CN113780012B (zh) * 2021-09-30 2023-12-29 东南大学 基于预训练语言模型的抑郁症访谈对话生成方法
CN113780012A (zh) * 2021-09-30 2021-12-10 东南大学 基于预训练语言模型的抑郁症访谈对话生成方法
CN113837766A (zh) * 2021-10-08 2021-12-24 支付宝(杭州)信息技术有限公司 风险识别方法、装置和电子设备
CN114880472A (zh) * 2022-04-28 2022-08-09 支付宝(杭州)信息技术有限公司 数据处理方法、装置及设备
CN115033676B (zh) * 2022-06-22 2024-04-26 支付宝(杭州)信息技术有限公司 一种意图识别模型训练及用户意图识别方法及装置
CN115033676A (zh) * 2022-06-22 2022-09-09 支付宝(杭州)信息技术有限公司 一种意图识别模型训练及用户意图识别方法及装置
CN115840808A (zh) * 2022-12-27 2023-03-24 广州汉申科技中介服务有限公司 科技项目咨询方法、装置、服务器及计算机可读存储介质
CN115840808B (zh) * 2022-12-27 2023-08-11 广州汉申科技中介服务有限公司 科技项目咨询方法、装置、服务器及计算机可读存储介质
CN116307663A (zh) * 2023-02-16 2023-06-23 广州力麒智能科技有限公司 一种智能自助政务服务系统、政务服务通办的方法及终端
CN116307663B (zh) * 2023-02-16 2023-09-08 广州力麒智能科技有限公司 一种智能自助政务服务系统、政务服务通办的方法及终端
CN116756277A (zh) * 2023-04-20 2023-09-15 海尔优家智能科技(北京)有限公司 基于目标生成式预训练gpt模型的交互语句的处理方法
CN116756277B (zh) * 2023-04-20 2023-11-24 海尔优家智能科技(北京)有限公司 基于目标生成式预训练gpt模型的交互语句的处理方法
CN116303982A (zh) * 2023-05-24 2023-06-23 杭州实在智能科技有限公司 基于rpa和自学习机制的智能应答与业务处理方法及系统
CN116303982B (zh) * 2023-05-24 2023-08-22 杭州实在智能科技有限公司 基于rpa和自学习机制的智能应答与业务处理方法及系统
CN117252211A (zh) * 2023-10-18 2023-12-19 中建八局第一数字科技有限公司 提问意图识别模型的构建、提问意图识别方法及装置

Also Published As

Publication number Publication date
CN108427722A (zh) 2018-08-21

Similar Documents

Publication Publication Date Title
WO2019153522A1 (zh) 智能交互方法、电子装置及存储介质
Poongodi et al. Chat-bot-based natural language interface for blogs and information networks
US11663411B2 (en) Ontology expansion using entity-association rules and abstract relations
US10665226B2 (en) System and method for data-driven socially customized models for language generation
US11397762B2 (en) Automatically generating natural language responses to users' questions
JP6667504B2 (ja) オーファン発話検出システム及び方法
Binali et al. Computational approaches for emotion detection in text
Sidorov et al. Empirical study of machine learning based approach for opinion mining in tweets
US20160364377A1 (en) Language Processing And Knowledge Building System
EP3714380A1 (en) Providing a summary of a multimedia document in a session
Arshad et al. Corpus for emotion detection on roman urdu
CN110597952A (zh) 信息处理方法、服务器及计算机存储介质
AU2011274318A1 (en) System and method of providing a computer-generated response
US20220405484A1 (en) Methods for Reinforcement Document Transformer for Multimodal Conversations and Devices Thereof
Wu et al. Posting act tagging using transformation-based learning
Kavitha et al. Chatbot for healthcare system using Artificial Intelligence
Dey et al. Emotion extraction from real time chat messenger
Sahu et al. Sarcasm detection: a review, synthesis and future research agenda
Kumar et al. An Ensemble-Based Model for Sentiment Analysis of Dravidian Code-Mixed Social Media Posts.
Nguyen et al. Building a chatbot system to analyze opinions of english comments
Lin et al. An emotion recognition mechanism based on the combination of mutual information and semantic clues
Rolih Applying coreference resolution for usage in dialog systems
US11386273B2 (en) System and method for negation aware sentiment detection
Lee Natural Language Processing: A Textbook with Python Implementation
CN112905752A (zh) 一种智能交互方法、装置、设备以及存储介质

Legal Events

Date Code Title Description
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 1205 DATED 08/12/2020)

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

Ref document number: 18905698

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 18905698

Country of ref document: EP

Kind code of ref document: A1