CN111046132B - Customer service question-answering processing method and system for searching multiple rounds of conversations - Google Patents
Customer service question-answering processing method and system for searching multiple rounds of conversations Download PDFInfo
- Publication number
- CN111046132B CN111046132B CN201911022722.6A CN201911022722A CN111046132B CN 111046132 B CN111046132 B CN 111046132B CN 201911022722 A CN201911022722 A CN 201911022722A CN 111046132 B CN111046132 B CN 111046132B
- Authority
- CN
- China
- Prior art keywords
- information
- service
- current information
- user
- domain
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
- G06F16/3329—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A customer service question-answering processing method and a system thereof for searching multiple rounds of conversations, wherein the customer service question-answering processing method comprises the following steps: acquiring current information input by a user; identifying the attributive service field according to the current information, screening a plurality of candidate reply messages from the service field, and extracting historical information which is input by a user and is related to the service field; searching and matching each piece of candidate reply information with context information formed by the current information and the historical information to obtain a corresponding matching degree score; and determining candidate reply information corresponding to the highest matching degree score as reply content of the current information. When the reply content of the current information is determined according to the service field, each piece of candidate reply information is searched and matched with the context information formed by the current information and the historical information, so that the limitation problem of searching and matching only with the current information can be avoided, the real intention of the user can be accurately understood from the context, and the accuracy of searching and matching is improved.
Description
Technical Field
The invention relates to the technical field of information processing, in particular to a customer service question-answering dialogue processing method and a system for retrieving multiple rounds of dialogues.
Background
The intelligent customer service system is an industry-oriented technical industry developed on the basis of large-scale knowledge processing, is applicable to the technical industries of large-scale knowledge processing, natural language understanding, knowledge management, automatic question-answering, reasoning and the like, provides fine-granularity knowledge management technology for enterprises, establishes a quick and effective technical means based on natural language for communication between the enterprises and massive users, and simultaneously can provide statistical analysis information required by the fine management for the enterprises.
At present, intelligent customer service plays an increasingly important role in the insurance industry, can replace manual work in a plurality of links such as consultation before insurance sale, service guidance, shopping guide, health check, rate calculation, after-sale check and claim, and the like, improves service timeliness, reduces operation cost, and is being integrated with more AI technologies to solve more basic repeatability problems.
However, in the conventional intelligent customer service multi-round dialogue interaction process, user problem sentences are simply connected, so that some disadvantages exist: 1) The interactive relation between the historical round problems is ignored; 2) The context has the lengthy and redundant situation, which is equivalent to introducing much noise, resulting in lower reply accuracy. Due to the adverse effects caused by the defects, the intelligent customer service system cannot accurately acquire the user intention, and the efficiency and experience of acquiring information of the user are affected.
Disclosure of Invention
The invention mainly solves the technical problem of how to improve the response accuracy and the user experience of the existing intelligent customer service system.
According to a first aspect, in one embodiment, a customer service question-answering processing method for retrieving multiple rounds of conversations is provided, including the following steps: acquiring current information input by a user; identifying the attributive service field according to the current information, screening a plurality of candidate reply messages from the service field, and extracting the history information which is input by the user and is related to the service field; searching and matching each piece of candidate reply information with the context information formed by the current information and the historical information to obtain a corresponding matching degree score; and determining the candidate reply information corresponding to the highest matching degree score as the reply content of the current information.
The identifying the attributive service field according to the current information comprises the following steps: performing text coding on the current information to obtain text vector representation; carrying out feature induction on the text vector representation to obtain the classification features of the current information, and identifying one or more fields and respective vector representations according to the classification features of the current information; respectively carrying out relation measurement on the text vector representation and the vector representation of each field to obtain probability values classified into each field; and determining the service domain to which the current information should belong from the domains according to the probability values.
The determining, from the domains according to the probability values, a service domain to which the current information should belong, including: judging that the probability value classified into each domain is larger than or equal to a first preset threshold value of the domain, and setting the domain as a service to be determined; when traversing each field and setting one or more to-be-determined services, detecting whether each to-be-determined service belongs to a preset service related field, if so, identifying the current information as a service field to which the current information should belong until detecting that each to-be-determined service is completed; and when the services to be determined do not belong to the fields related to the preset services, sending chatting information to the user, wherein the chatting information comprises randomly formed recommended contents.
If the fields are traversed and the service to be determined is not set, judging whether the probability value classified into each field is larger than or equal to a second preset threshold value of the field, if so, setting the field as the service to be clarified, and if not, setting the field as the refused service; the second preset threshold value is smaller than the first preset threshold value; when traversing each domain and setting one or more to-be-clarified services, sending out problem clarification information to the user so that the user confirms at least one domain from each to-be-clarified service; and when the fields are traversed and no service to be clarified is set, switching the user to the manual customer service.
And searching and matching each piece of candidate reply information with the context information formed by the current information and the historical information to obtain a corresponding matching degree score, wherein the searching and matching comprises the following steps: forming the current information and the historical information into context information, wherein the context information comprises at least two context questions input by the user and related to the service field; text analysis is carried out on each context problem by utilizing a text processing model with multiple granularities, and text analysis is carried out on each candidate reply message, so that multiple granularities of representations corresponding to the context problem and multiple granularities of representations corresponding to the candidate reply message are obtained; performing interactive processing on various granularity representations corresponding to the context problems and various granularity representations corresponding to each piece of candidate reply information, and calculating to obtain an interactive representation sequence corresponding to the piece of candidate reply information; and carrying out vector matching processing on the interactive representation sequence according to a preset cyclic neural network, and outputting a corresponding matching degree score.
The plurality of granularity text processing models includes at least two of the following models: a Word vector representation model, a Word2Vec Word vector representation model, a local contextualized semantic vector representation model, a serial contextualized semantic vector representation model, a self-attention mechanism semantic vector representation model and an interactive attention semantic vector representation model based on a convolutional neural network.
Performing interaction processing on the multiple granularity representations corresponding to the context questions and the multiple granularity representations corresponding to each piece of candidate reply information, and calculating to obtain an interaction representation sequence corresponding to the piece of candidate reply information, wherein the interaction representation sequence comprises the following steps: if each of the context questions is defined as U i Each candidate reply message is R r I and r are context problem sequence numbers and candidate reply sequence numbers respectively; calculating R r Concerning U i Is formulated as a weighted representation of
Wherein the method comprises the steps of，
Representing U i Is>
The pair represents R r K is the granularity representation number, exp () is an exponential function, tanh () is a hyperbolic tangent function, n r R represents r The total word number of the corresponding r candidate reply message,/->
W a 、b a All represent super parameters; for->
And->
Performing interactive operation to obtain
Wherein f () represents an interactive function, reLU () represents an active function, W p 、b p All represent super parameters; according to t i,j Obtaining R r Corresponding interactive representation sequence
Wherein n is i Representing U i Is a total number of columns.
Vector matching processing is carried out on the interactive representation sequence according to a preset cyclic neural network, and a corresponding matching degree score is output, and the method comprises the following steps: using a preset cyclic neural network to sequence the interactive representation 
To be generalized as a first vector V i Obtain a matching vector list (V 1 ,...,V i ,....,V m ) Wherein m is the total number of the context questions; reusing the recurrent neural network to list the matching vectors (V 1 ,...,V i ,....,V m ) Generalized to second vector C r The second vector C r And after passing through a linear layer, outputting the matching degree score corresponding to the r-th candidate reply message until the matching degree score corresponding to each candidate reply message is obtained.
According to a second aspect, in one embodiment, there is provided a customer service questioning and answering processing system, comprising: the communication service module is used for receiving consultation information input by a user, wherein the consultation information comprises current information and historical information input by the user; the question-answer processing module is connected with the communication service module and is used for determining the reply content of the current information according to the customer service question-answer processing method in the first aspect; and the response reply module is connected with the question-answer processing module and is used for feeding back the determined reply content to the user.
According to a third aspect, there is provided in one embodiment a computer-readable storage medium including a program executable by a processor to implement the customer service questioning and answering processing method described in the above third aspect.
The beneficial effects of this application are:
according to the customer service question-answering processing method and the system thereof for searching multiple rounds of conversations, the customer service question-answering processing method comprises the following steps: acquiring current information input by a user; identifying the attributive service field according to the current information, screening a plurality of candidate reply messages from the service field, and extracting historical information which is input by a user and is related to the service field; searching and matching each piece of candidate reply information with context information formed by the current information and the historical information to obtain a corresponding matching degree score; and determining candidate reply information corresponding to the highest matching degree score as reply content of the current information. In the first aspect, when the service domain to which the current information belongs is identified, the service domain to which the current information should belong is determined from each domain according to the probability value classified into each domain, so that the identification accuracy of the service domain can be improved, and the attention problem of the user can be accurately replied; in the second aspect, in the process of identifying the current information, the application not only can determine the service field to which the current information belongs according to the identification result, but also takes the fact that the field does not meet the service requirement and is boring and replying, the field is unclear and clarifying and replying, and the possibility that the field cannot be identified and is switched to the artificial customer service is considered, so that the implementation mode of customer service question and answer is enriched, the consultation needs of the user are met to the greatest extent, and the consultation experience of the user is promoted; in the third aspect, when determining the reply content of the current information according to the service field, each piece of candidate reply information is searched and matched with the context information formed by the current information and the historical information, so that the limitation problem of searching and matching only with the current information can be avoided, the real intention of the user can be accurately understood from the context of the context, and the accuracy of searching and matching is improved; in the fourth aspect, when calculating the matching degree score corresponding to each candidate reply message, text analysis is performed on each context problem and each candidate reply message respectively by using text processing models with a plurality of granularities, so that the matching degree score is obtained through interactive processing and vector matching processing, the advantages of different text processing models can be fully utilized, the reliability of the matching degree score is enhanced, and the candidate reply message with highest reliability is determined from the text processing models; in a fifth aspect, the customer service question-answering processing system constructed by the communication service module, the question-answering processing module and the response reply module can accurately search and match the context information and the reply content by using a multi-granularity text representation method, so that the natural and smooth multi-round interactive question-answering requirement is realized, the efficiency of the intelligent customer service dialogue system can be effectively improved, higher-standard consultation experience is provided for users, and the customer service question-answering processing system has great practical value.
Drawings
FIG. 1 is a flow chart of a customer service questioning and answering processing method for retrieving multiple rounds of conversations in the present application;
FIG. 2 is a flow chart for identifying a home business segment based on current information;
FIG. 3 is a flow chart of determining customer service question-answering modes under different situations according to probability values categorized into various fields;
FIG. 4 is a flow chart of obtaining a corresponding match score by retrieving matches;
FIG. 5 is a schematic diagram of a domain identification model;
FIG. 6 is a schematic diagram of a principle of calculating a matching score;
FIG. 7 is a schematic diagram of a customer service question-answering processing method;
fig. 8 is a schematic structural diagram of a customer service answering processing system in the present application.
Detailed Description
The invention will be described in further detail below with reference to the drawings by means of specific embodiments. Wherein like elements in different embodiments are numbered alike in association. In the following embodiments, numerous specific details are set forth in order to provide a better understanding of the present application. However, one skilled in the art will readily recognize that some of the features may be omitted, or replaced by other elements, materials, or methods in different situations. In some instances, some operations associated with the present application have not been shown or described in the specification to avoid obscuring the core portions of the present application, and may not be necessary for a person skilled in the art to describe in detail the relevant operations based on the description herein and the general knowledge of one skilled in the art.
Furthermore, the described features, operations, or characteristics of the description may be combined in any suitable manner in various embodiments. Also, various steps or acts in the method descriptions may be interchanged or modified in a manner apparent to those of ordinary skill in the art. Thus, the various orders in the description and drawings are for clarity of description of only certain embodiments, and are not meant to be required orders unless otherwise indicated.
The numbering of the components itself, e.g. "first", "second", etc., is used herein merely to distinguish between the described objects and does not have any sequential or technical meaning. The terms "coupled" and "connected," as used herein, are intended to encompass both direct and indirect coupling (coupling), unless otherwise indicated.
For a clear understanding of the technical solutions of the present application, some technical terms will be explained here.
Text classification algorithms are roughly divided into two categories, namely ignoring word order and performing shallow semantic modeling on text (a representative model comprises LDA, earth mole's distance); one is to consider word order, and perform deep semantic modeling on text (a representative model has a deep learning algorithm, such as LSTM, CNN, etc.). In the deep learning model, spatial patterns (spatial patterns) are summarized at a lower level to help represent higher level concepts, e.g., CNN build convolution feature detectors extract patterns from local sequence windows and use max pooling to select obvious features.
Convolutional neural network, (Convolutional Neural Network, CNN) is a feed-forward neural network whose artificial neurons can respond to surrounding cells in a part of the coverage area with excellent performance for large image processing. Convolutional neural networks consist of one or more convolutional layers and a top fully connected layer (corresponding to classical neural networks) and also include associated weights and pooling layers (pooling layers). This structure enables the convolutional neural network to take advantage of the two-dimensional structure of the input data. Convolutional neural networks can give better results in terms of image and speech recognition than other deep learning structures. This model may also be trained using a back propagation algorithm. In addition, convolutional neural networks require fewer parameters to consider, making them an attractive deep learning architecture.
The gating cyclic neural network introduces the concepts of a reset gate (reset gate) and an update gate (update gate), so that the calculation mode of the hidden state in the cyclic neural network is modified, the flow of information is controlled through a gate which can be learned, and the dependency relationship with larger time step distance in a time sequence is better captured.
The capsule network belongs to a special neural network, and is a brand new neural network which is trained in a Dynamic Routing mode by using a neuron vector to replace a single neuron node of the traditional neural network. Because CNN models spatial information, the feature detectors need to be duplicated, and the number of feature detectors to be duplicated or the number of label data to be duplicated increases exponentially with the data dimension, reducing the model efficiency. The prior neural network space insensitive method is inevitably limited by rich text structures (such as the position information of the saved words, semantic information, grammar structures and the like), while the capsule network effectively improves the defects of the CNN neural network method.
The transducer network uses an attribute mechanism to reduce the distance between any two positions in the sequence to a constant; secondly, the method is not similar to the sequential structure of the cyclic neural network, so that the method has better parallelism and accords with the existing GPU framework.
Word2Vec Word vectors, which are feature vectors or tokens of words, map words into real number domain vectors.
The question-answering dialogue system comprises an NLU (natural language understanding process) and an NLG (natural language generation) process. The NLU process is to convert the human language into a machine understood language, and the NLG process is to give an answer in a natural language generation mode after processing through a dialogue system.
The technical scheme of the present application will be described below with reference to examples.
Embodiment 1,
Referring to fig. 1, the present application discloses a customer service inquiry processing method for retrieving multiple rounds of conversations, which includes steps S100-S400, and is described below.
Step S100, current information input by a user is acquired.
The user can input any information, usually a consultation problem, at the mobile terminal, the computer, the service robot and other equipment, the information is transmitted to the background server with the intelligent customer service question-answering function, and the background server replies after acquiring the information, so that the dialogue between the user and the background server is started. In the multi-turn conversation process of the two, the information currently input by the user can be called current information, and the information already input by the user and replied to by the server can be called history information.
Step S200, identifying the attributive service domain according to the current information, screening a plurality of candidate reply messages from the service domain, and extracting the history information related to the service domain, which is input by the user.
Before replying to the current information input by the user, the background server with the intelligent customer service question-answering function needs to accurately determine the consultation intention of the user from the current information, and then needs to identify the knowledge field pointed by the consultation problem by utilizing the information such as key words, context, semantics and the like contained in the current information, so that the consultation problem is attributed to the service field supported by the background server, and the replying is performed according to the service knowledge stored by the background server.
And step S300, searching and matching each piece of candidate reply information with the context information formed by the current information and the historical information to obtain a corresponding matching degree score.
In this embodiment, the current information and the history information may be combined to form the context information, and the context information may better focus on the real intention of the user through context and semantics, so that when the context information is used to search and match with the candidate reply information, a more accurate matching degree score may be obtained.
Step S400, determining candidate reply information corresponding to the highest matching degree score as the reply content of the current information. The matching degree score is a quantitative index for evaluating the degree of correlation of the message, and the higher the score is, the higher the degree of correlation between the reply content and the consultation problem is, and when the candidate reply information corresponding to the highest matching degree score is used as the reply content, the accurate reply can be performed aiming at the consultation intention of the user, and the question and answer experience of the user is improved.
In this embodiment, referring to fig. 2, the above-mentioned step S200 involves a process of identifying the service domain to which the service belongs according to the current information, and the process may be specifically described through steps S210 to S240, which are respectively described below.
Step S210, text encoding is carried out on the current information, and text vector representation is obtained.
It should be noted that, devices such as a computer and a server cannot directly process natural language, and need to perform text encoding on information, so as to process the information into a text vector representation in a digital form and capable of being identified.
And step S220, carrying out feature induction on the text vector representation to obtain the classification features of the current information, and identifying one or more fields and respective vector representations according to the classification features of the current information.
It should be noted that the categorized features obtained by feature induction may be digital information such as key words, context, and semantics, and represent knowledge in different fields, so that one or more fields to which the feature induction belongs may be identified by the knowledge. Similarly, in computers, servers, and the like, knowledge and domains are recorded using vector representations in digital form.
In step S230, the text vector representation and the vector representation in each domain are respectively subjected to a relationship measurement, so as to obtain probability values categorized into each domain. The relationship measure is a calculation process for evaluating the degree of correlation, and the evaluation result is often expressed as a probability value.
In one embodiment, see FIG. 5, a domain identification model may be built to identify and categorize the current information. For example, a text classification model based on a capsule network is adopted to process the current information input by a user, the text classification model is obtained by training and learning text classification training corpus, the model mainly comprises three processing modules of a coding layer, a feature induction layer and a relation measurement layer, and vector representation of each field category is obtained besides the optimal model parameters obtained by training. In the process of using a text classification model based on a capsule network for practical application, information input by a user is encoded into text vector representations through an encoder, and compared with the vector representations in each field in a knowledge base, the similarity is calculated, so that probability values classified into each field are obtained.
And step S240, determining the service domain to which the current information should belong from the domains according to the probability values classified into the domains. In one particular embodiment, referring to FIG. 3, step 240 may include steps S241-S249 and steps S251-S253, respectively, as described below.
Step S241, judging whether the probability value classified into each domain is greater than or equal to the first preset threshold value of the domain, if yes, proceeding to step S242, otherwise proceeding to step S247.
The first preset threshold is a manually set identification standard, and when the first preset threshold is greater than or equal to the first preset threshold in the field, the true meaning of the current information input by the user is indicated to be clear, and the current information can be understood by a machine and divided into a specific knowledge field.
In step S242, the domain corresponding to the probability value is set as the service to be determined. In addition, after traversing the domain identified in step S220 according to the method of step S241, one or more services to be determined may be provided.
Step S243 is to detect whether each service to be determined belongs to the preset service related field, if so, step S244 is entered, otherwise step S246 is entered.
The background server can only reply to the content related to the service scope supported by the background server, so that it is required to further confirm whether the service to be determined belongs to the supported service field (such as the insurance service in the insurance industry, the legal service, the infringement of the lawyer industry, or the service in other industries) or belongs to the non-related service field.
Step S244, the current information is identified as the service area to which the current information should belong. In addition, according to the method of step S243, each service to be determined is traversed until each service to be determined is detected to be completed, and one or more service areas to which the service should belong can be obtained.
Step S245, a plurality of candidate reply messages are screened from the determined one or more service areas. Specifically, a plurality of candidate reply messages may be screened out in each service domain, thereby forming a plurality of candidate reply messages herein.
Step S246, when each service to be determined does not belong to the field related to the preset service, the method indicates that the consultation problem proposed by the user exceeds the service processing range of the background server, at the moment, the chatting information can be sent to the user, the chatting information can comprise randomly formed recommended content, and in this way, the user can notice that the problem exceeds the range of the consultation service, and the embarrassing situation that nothing is replied can be avoided.
Step S247, which is entered when it is determined that the probability value categorized into each domain is smaller than the first preset threshold value of the domain. The respective domains may be traversed according to the method of step S241, and a case may occur in which the probability value of each domain is smaller than the first preset threshold value of the domain, where no service to be determined is considered to be set.
Step S248, judging whether the probability value classified into each domain is greater than or equal to the second preset threshold value of the domain, if yes, proceeding to step S249, otherwise proceeding to step S252.
The second preset threshold is also a manually set identification standard, and when the second preset threshold is greater than or equal to the second preset threshold in the field, the current information input by the user can be understood by a machine, but a certain ambiguity exists, which may be that the user mixes a plurality of intentions into a sentence or the user lays a long time before expressing the intentions, and the situations need to be clarified further with the user.
Step S249, the domain corresponding to the probability value is set as the service to be clarified. One or more businesses to be clarified may be set by traversing each domain according to the method of step S248, that is, the user' S consultation problem includes multiple knowledge domains at the same time, and verification is required for these knowledge domains.
Step S251, when traversing each domain and setting one or more to-be-clarified services, issue clarification information can be sent to the user so that the user can confirm at least one domain from each to-be-clarified service, and thus the background server can clearly know the real intention of the user.
Step S252, the step is entered when judging that the probability value classified to each domain is smaller than the second preset threshold value of the domain. The method of step S248 may traverse each domain, and the probability value of each domain may be smaller than the second preset threshold value of the domain, where the service to be clarified may not be set, that is, the proposed consultation problem of the user is in a refused state, and the background server cannot learn the true intention and the fuzzy intention of the user at all, so that the machine cannot accept the consultation problem of the user any more, and step S253 may be performed.
In step S253, the user is transferred to the manual customer service, and the customer service personnel communicates with the user, so that the user expression can be better understood by the communication between people.
In this embodiment, referring to fig. 4, the process of calculating the matching score corresponding to each candidate reply message in the step S300 described above may be specifically described through steps S310 to S340, which are respectively described below.
In step S310, the current information and the history information are formed into context information, where the context information includes at least two context questions related to the business segment inputted by the user.
It will be appreciated that the information entered by the user each time in the current session or in the history session can be considered as a problem of a consulting nature, and then in many rounds of sessions it will be a problem of a plurality of consulting nature. If the current information is related to one or more business fields to which it belongs, there may be one or more questions in the history dialogue related to the business fields, and the questions related to the business fields are marked, so that at least two questions in the same context can be obtained, and each question in the same context may be referred to as a context question.
In step S320, text analysis is performed on each context question using a text processing model with multiple granularities, and text analysis is performed on each candidate reply message, so as to obtain multiple granularities of representations corresponding to the context question and multiple granularities of representations corresponding to the candidate reply message.
In this embodiment, the plurality of granularity text processing models includes at least two of the following models: a Word vector representation model, a Word2Vec Word vector representation model, a local contextualized semantic vector representation model, a serial contextualized semantic vector representation model, a self-attention mechanism semantic vector representation model and an interactive attention semantic vector representation model based on a convolutional neural network.
A brief description of each granularity of the text processing model will be provided herein. 1) A word vector representation model based on a convolutional neural network is characterized in that a text word in a user problem is cut into words, the words are converted into word vector representations through the word vector model, and the word vector representations are obtained through the convolutional neural network and the maximum pooling operation. 2) Word2Vec Word vector representation model is co-occurrence information of the code Word in the corpus. 3) The local contextualized semantic vector representation model is used for carrying out Word2Vec Word vector representation on user question sentences and modeling local dependence among words by applying a convolutional neural network and maximum pooling operation. 4) The sequence contextualized semantic vector representation model is used for carrying out Word2Vec Word vectorization on user questions and capturing sequence relations among words by applying gating cyclic neural network operation. 5) The semantic vector representation model of the Self-attention mechanism is characterized in that Word2Vec Word vectorization is carried out on a user question, self-attention operation is carried out by using a transducer network, K and V are input into the user question, each Word in the sentence can be represented by a similar Word in the sentence, dependence among remote words in the sentence can be encoded, and global dependence information of words in the sentence is modeled. 6) The interactive attention mechanism semantic vector representation model respectively carries out Word2Vec Word vector representation on user question and candidate reply, a trans-former network is applied to carry out Cross-attention operation, and input K and V are respectively the user question and the candidate reply, so that words in the user question can be represented by similar words in the candidate reply, direct connection between the user question and the candidate reply can be established, and key information required by matching of the questions and the replies is extracted.
It will be appreciated by those skilled in the art that the text vector representation models with each granularity mentioned above are all text processing tools commonly used by technicians, and all belong to the prior art, and can be known from the technical literature disclosed so far, and for avoiding redundancy, the working principles and application modes of the text vector representation models will not be limited and specifically described herein.
Step S330, performing interaction processing on the multiple granularity representations corresponding to the context questions and the multiple granularity representations corresponding to each piece of candidate reply information, and calculating to obtain an interaction representation sequence corresponding to the candidate reply information.
In one embodiment, if each context question is defined as U i Each candidate reply message is R r And i and r are context problem sequence numbers and candidate reply sequence numbers respectively. Then R can be calculated r Concerning U i Is formulated as a weighted representation of
Wherein, the liquid crystal display device comprises a liquid crystal display device,
representing U i Is>
The pair represents R r K is the granularity representation number, exp () is an exponential function, tanh () is a hyperbolic tangent function, n r R represents r The total word number of the corresponding r candidate reply message,/->
W a 、b a All represent hyper-parameters.
Further, to 
And->
Performing interactive operation to obtain
Wherein f () represents an interactive function, reLU () represents an active function, W p 、b p All represent hyper-parameters.
Further, according to t i,j Obtaining R r Corresponding interactive representation sequence
Wherein n is i Representing U i Is a total number of columns.
And step S340, carrying out vector matching processing on the interactive representation sequence according to a preset cyclic neural network, and outputting a corresponding matching degree score.
In one embodiment, the interactive representation sequence may be performed using a predetermined recurrent neural network
To be generalized as a first vector V i Obtain a matching vector list (V 1 ,...,V i ,....,V m ) Where m is the total number of the context questions.
Further, the cyclic neural network is reused to list the matching vectors (V 1 ,...,V i ,....,V m ) Generalized to second vector C r Will second vector C r After passing through a linear layer, outputting the matching degree score corresponding to the r candidate reply message until the matching degree score corresponding to each candidate reply message is obtained.
It should be noted that, the recurrent neural network mentioned in this embodiment may be a gated recurrent neural network (Gated Recurrent Unit, GRU), and the GRU introduces the concepts of a reset gate (reset gate) and an update gate (update gate), so as to modify the previous calculation method of the hidden state in the recurrent neural meridian, and control the flow of information through a gate that can be learned, so as to better capture the dependency relationship with a larger time step distance in the time sequence. The input to the GRU is a sequence, each step of the sequence is a vector, and the output is typically the hidden state vector of the last step of the hidden layer.
It should be noted that, the Linear layer (Linear layer) in this embodiment is a common structural layer of the neural network, and the Linear layer of the neural network may implement an operation of affine transformation on data, and may be a function of combining features. Since both the gated recurrent neural network and the linear layer are within the scope of the prior art, they will not be described in detail herein.
In one implementation, referring to fig. 6, the steps S310-S340 may be summarized as a text representation phase and a candidate reply matching phase, respectively, as follows.
In the text representation phase, each context question may be represented as a context question U 1 To context problem U m Respectively carrying out text analysis on the context problems through 6 text vector representation models with granularity to obtain a context problem U 1 Corresponding 6 granularity representations, context problem U i The corresponding 6 granularity representations, and so on. Then, for one candidate reply message R of the candidate reply messages r The candidate reply information R is subjected to text vector representation model with 6 granularity r Text analysis is carried out to obtain candidate reply information R r The corresponding 6 particle sizes are shown.
In the candidate reply matching phase, each context problem (U 1 To U (U) m ) Corresponding multiple granularity representation and strip candidate reply information R r Performing interactive processing on the corresponding multiple granularity representations, thereby calculating and obtaining the candidate reply information R r Corresponding interactive representation sequence
Interactive presentation sequences can be +.>
To be generalized as a first vector V i Obtain a matching vector list (V 1 ,...,V i ,....,V m ) And then the matching vector list (V) is formed by using the cyclic neural network GRU 1 ,...,V i ,....,V m ) Generalized to second vector C r Will second vector C r After passing through the linear layer, outputting the R candidate reply message R r The corresponding match score.
It will be appreciated that each candidate reply message (R) may be processed through a text representation phase and a candidate reply matching phase 1 、…，R r ，…，R n Where n represents the total number of selected reply messages) to the matching degree score.
In order to make the technical solution of the present application clearly known to those skilled in the art, the working principle of the customer service inquiry and answer processing method in this embodiment is described herein according to fig. 6 and 7.
The client question-answering processing method can be divided into a user input process, an NLU (natural language understanding) parsing process, an NLU (natural language understanding) output process, an NLG (natural language generation) process, a library checking process, and a reply output process.
In the process of user input, the current information and the history information input by the user can be obtained through multiple rounds of conversations.
In the NLU parsing process, language parsing can be performed on the current information and the history information through the domain recognition model introduced in step S230, so as to obtain probability values of classifying the current information into each domain (the history information processing process may refer to the current information, and details are not repeated here). Three recognition results can be obtained through step S240: the domain cannot be determined (i.e. the case where no traffic to be clarified is set), the result is to be clarified (i.e. the case where one or more traffic to be clarified is set), the domain can be determined (i.e. the case where one or more traffic to be determined is set).
In the NLU output process, a user can be switched to manual customer service when the field cannot be determined, problem clarification processing is performed when the result is clarified, and whether the result belongs to the field related to the preset service or not is further judged when the field can be determined.
In the NLG process, when the domain can be determined, the results of the domain which does not belong to the service domain (namely, the case of the domain which belongs to the preset service) and the result which belongs to the service domain (namely, the case of the domain which belongs to the preset service) can be further obtained, the chatting reply processing is carried out when the domain which does not belong to the service domain is obtained, and the retrieval matching processing is carried out when the domain which belongs to the service domain is obtained.
In the process of checking the database, the corresponding log database is accessed based on the service field to which the current information belongs, a plurality of candidate reply messages are screened out from the log database and stored in a candidate pool, and the candidate reply messages are waited for further searching and matching processing, so that the candidate reply information corresponding to the highest matching degree score is determined as the reply content of the current information.
In the reply output process, under the condition of the chat reply processing, sending the chat reply content to the user; in the case of the search matching process, candidate reply information corresponding to the highest matching degree score is transmitted to the user.
Those skilled in the art will appreciate that the following application advantages may be achieved by the steps S100-S400 described above: (1) When the service domain to which the current information belongs is identified, the service domain to which the current information should belong is determined from each domain according to the probability value classified into each domain, so that the identification accuracy of the service domain can be improved, and the attention problem of a user can be accurately recovered; (2) In the process of identifying the current information, not only the service field to which the service should belong can be determined according to the identification result, but also the boring reply is carried out in consideration of the fact that the field does not accord with the service requirement, the clear reply is carried out because the field is unclear, and the possibility that the field cannot be identified and is switched to the artificial customer service is enriched, the implementation mode of customer service inquiry and reply is enriched, the consultation needs of the user are met to the maximum extent, and the consultation experience of the user is promoted; (3) When the reply content of the current information is determined according to the service field, each piece of candidate reply information is searched and matched with the context information formed by the current information and the historical information, so that the limitation problem of searching and matching only with the current information can be avoided, the real intention of a user can be accurately understood from the context of the context, and the accuracy of searching and matching is improved; (4) When the matching degree score corresponding to each candidate reply message is calculated, text analysis is carried out on each context problem and each candidate reply message by using a plurality of text processing models with granularity, so that the matching degree score is obtained through interactive processing and vector matching processing, the advantages of different text processing models can be fully utilized, the credibility of the matching degree score is enhanced, and the candidate reply message with highest credibility is determined from the text processing models.
Embodiment II,
Referring to fig. 8, on the basis of the customer service question-answering processing method disclosed in the first embodiment, the present application further discloses a customer service question-answering processing system 1, which includes a communication service module 11, a question-answering processing module 12 and a response reply module 13, and is described below.
The communication service module 11 is configured to receive advisory information input by a user, where the advisory information includes current information and history information input by the user.
The communication service module 11 may be an information input unit on a mobile terminal, a computer, or a service robot, and the user may input any information through the information input unit. The information input by the user is usually a consultation problem, and the information is transmitted to a background server with an intelligent customer service question-answering function, and the background server replies after acquiring the information, so that a dialogue between the user and the background server is started. In the multi-turn conversation process of the two, the information currently input by the user can be called current information, and the information already input by the user and replied to by the server can be called history information.
The question-answer processing module 12 is connected to the communication service module 11, which may be a data processing unit on a background server, and determines the reply content of the current information mainly according to the customer service question-answer processing method disclosed in the first embodiment, and may select a reply mechanism for entering into a manual customer service, a reply mechanism for clarifying a problem, a reply mechanism for boring a reply process, and a reply mechanism for retrieving a matching process according to the domain identification result of the current information. The specific function of the question-answering processing module 12 may refer to the first embodiment, and will not be described here.
The response reply module 13 is connected with the question and answer processing module 12 and the communication service module 11, and may include a log library related to the service domain, may output some candidate reply messages in response to the query of the question and answer processing module, and may timely send the candidate reply messages and the boring reply messages to the communication service module 11 for the user to review.
It can be understood by those skilled in the art that the customer service question-answering processing system constructed by the communication service module, the question-answering processing module and the response reply module can accurately search and match the context information and the reply content by using a multi-granularity text representation method, thereby realizing natural and smooth multi-round interactive question-answering requirements, effectively improving the efficiency of the intelligent customer service dialogue system, providing higher-standard consultation experience for users and having great practical value.
Those skilled in the art will appreciate that all or part of the functions of the various methods in the above embodiments may be implemented by hardware, or may be implemented by a computer program. When all or part of the functions in the above embodiments are implemented by means of a computer program, the program may be stored in a computer readable storage medium, and the storage medium may include: read-only memory, random access memory, magnetic disk, optical disk, hard disk, etc., and the program is executed by a computer to realize the above-mentioned functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the functions described above can be realized. In addition, when all or part of the functions in the above embodiments are implemented by means of a computer program, the program may be stored in a storage medium such as a server, another computer, a magnetic disk, an optical disk, a flash disk, or a removable hard disk, and the program in the above embodiments may be implemented by downloading or copying the program into a memory of a local device or updating a version of a system of the local device, and when the program in the memory is executed by a processor.
The foregoing description of the invention has been presented for purposes of illustration and description, and is not intended to be limiting. Several simple deductions, modifications or substitutions may also be made by a person skilled in the art to which the invention pertains, based on the idea of the invention.
Claims (8)
1. A customer service question-answering processing method for searching multiple rounds of conversations is characterized by comprising the following steps:
acquiring current information input by a user;
identifying the attributive service field according to the current information, screening a plurality of candidate reply messages from the service field, and extracting the history information which is input by the user and is related to the service field;
and searching and matching each piece of candidate reply information with the context information formed by the current information and the historical information to obtain a corresponding matching degree score, wherein the method comprises the following steps of:
composing said current information and said history information into context information, said context information comprising at least two context questions related to said business segment entered by said user,
text analysis is carried out on each context question by utilizing a text processing model with a plurality of granularities, and text analysis is carried out on each candidate reply message to obtain a plurality of granularities of representations corresponding to the context question and a plurality of granularities of representations corresponding to the candidate reply message, wherein each context question is defined as U i Each candidate reply message is R r Wherein i and r are context problem sequence numbers and candidate reply sequence numbers respectively, and define
Representing U i Is>
R represents r J, k are context problem granularity representation sequence numbers, candidate reply granularity representation sequence numbers,
performing interaction processing on multiple granularity representations corresponding to each context problem and multiple granularity representations corresponding to each candidate reply message, and calculating to obtain an interaction representation sequence corresponding to the candidate reply message, wherein the interaction representation sequence comprises the following steps:
calculating R r Concerning U i Is a weighted representation of (2)
Expressed as by the formula
Wherein exp () is an exponential function, tanh () is a hyperbolic tangent function, n r R represents r The total word number of the corresponding r candidate reply message,
W a 、b a all of which represent the super-parameters,
for a pair of
And->
Performing interactive operation to obtain
Wherein f () represents an interactive function, reLU () represents an active function, W p 、b p All represent super parameters according to t i,j Obtaining R r Corresponding interactive representation sequence
Wherein n is i Representing U i Is a total number of columns;
vector matching processing is carried out on the interactive representation sequence according to a preset cyclic neural network, and a corresponding matching degree score is output;
and determining the candidate reply information corresponding to the highest matching degree score as the reply content of the current information.
2. The customer service question-answering processing method according to claim 1, wherein the identifying the service area to which the customer belongs based on the current information includes:
performing text coding on the current information to obtain text vector representation;
carrying out feature induction on the text vector representation to obtain the classification features of the current information, and identifying one or more fields and respective vector representations according to the classification features of the current information;
respectively carrying out relation measurement on the text vector representation and the vector representation of each field to obtain probability values classified into each field;
and determining the service domain to which the current information should belong from the domains according to the probability values.
3. The customer service question-answering processing method according to claim 2, wherein the determining, from each of the areas, a service area to which the current information should belong according to the probability value, includes:
judging that the probability value classified into each domain is larger than or equal to a first preset threshold value of the domain, and setting the domain as a service to be determined;
when traversing each field and setting one or more to-be-determined services, detecting whether each to-be-determined service belongs to a preset service field, if so, identifying the current information as a service field to which the current information should belong until detecting that each to-be-determined service is completed;
And when the services to be determined do not belong to the preset service field, sending chatting information to the user, wherein the chatting information comprises randomly formed recommended content.
4. A customer service question-answering processing method according to claim 3, wherein if each domain is traversed and no service to be determined is set, judging whether the probability value of each domain is greater than or equal to a second preset threshold value of the domain, if so, setting the domain as a service to be clarified, and if not, setting the domain as a refused service; the second preset threshold value is smaller than the first preset threshold value;
when traversing each domain and setting one or more to-be-clarified services, sending out problem clarification information to the user so that the user confirms at least one domain from each to-be-clarified service;
and when the fields are traversed and no service to be clarified is set, switching the user to the manual customer service.
5. The customer service questioning and answering processing method according to claim 1, wherein the plurality of granularity text processing models include at least two of the following models: a Word vector representation model, a Word2Vec Word vector representation model, a local contextualized semantic vector representation model, a serial contextualized semantic vector representation model, a self-attention mechanism semantic vector representation model and an interactive attention semantic vector representation model based on a convolutional neural network.
6. The customer service question-answering processing method according to claim 1, wherein the vector matching processing is performed on the interactive representation sequence according to a preset cyclic neural network, and a corresponding matching degree score is output, and the method comprises the following steps:
using a preset cyclic neural network to sequence the interactive representation
To be generalized as a first vector V i Obtain a matching vector list (V 1 ,...,V i ,....,V m ) Wherein m is the total number of the context questions;
reusing the recurrent neural network to list the matching vectors (V 1 ,...,V i ,....,V m ) Generalized to second vector C r The second vector C r And outputting the matching degree score corresponding to the r-th candidate reply message after passing through a linear layer until the matching degree score corresponding to each candidate reply message is obtained.
7. A customer service questioning and answering processing system, comprising:
the communication service module is used for receiving consultation information input by a user, wherein the consultation information comprises current information and historical information input by the user;
a question-answer processing module, connected to the communication service module, for determining a reply content of the current information according to the customer service question-answer processing method of any one of claims 1 to 6;
and the response reply module is connected with the question-answer processing module and is used for feeding back the determined reply content to the user.
8. A computer-readable storage medium, comprising a program executable by a processor to implement the customer service questioning and answering processing method according to any one of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911022722.6A CN111046132B (en) | 2019-10-25 | 2019-10-25 | Customer service question-answering processing method and system for searching multiple rounds of conversations |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911022722.6A CN111046132B (en) | 2019-10-25 | 2019-10-25 | Customer service question-answering processing method and system for searching multiple rounds of conversations |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111046132A CN111046132A (en) | 2020-04-21 |
| CN111046132B true CN111046132B (en) | 2023-06-16 |
Family
ID=70231845
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911022722.6A Active CN111046132B (en) | 2019-10-25 | 2019-10-25 | Customer service question-answering processing method and system for searching multiple rounds of conversations |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111046132B (en) |
Families Citing this family (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111488448B (en) * | 2020-05-27 | 2023-06-20 | 支付宝(杭州)信息技术有限公司 | Method and device for generating machine reading annotation data |
| CN111966782B (en) | 2020-06-29 | 2023-12-12 | 百度在线网络技术（北京）有限公司 | Multi-round dialogue retrieval method and device, storage medium and electronic equipment |
| CN111753074B (en) * | 2020-06-30 | 2021-08-17 | 贝壳找房(北京)科技有限公司 | Method, device, medium and electronic equipment for realizing session |
| CN111858854B (en) * | 2020-07-20 | 2024-03-19 | 上海汽车集团股份有限公司 | Question-answer matching method and relevant device based on historical dialogue information |
| CN111625641B (en) * | 2020-07-30 | 2020-12-01 | 浙江大学 | Dialog intention recognition method and system based on multi-dimensional semantic interaction representation model |
| CN112000787B (en) * | 2020-08-17 | 2021-05-14 | 上海小鹏汽车科技有限公司 | Voice interaction method, server and voice interaction system |
| CN112308370B (en) * | 2020-09-16 | 2024-03-05 | 湘潭大学 | Automatic subjective question scoring method for thinking courses based on Transformer |
| CN112182159B (en) * | 2020-09-30 | 2023-07-07 | 中国人民大学 | Personalized search type dialogue method and system based on semantic representation |
| CN112036173A (en) * | 2020-11-09 | 2020-12-04 | 北京读我科技有限公司 | Method and system for processing telemarketing text |
| CN112565663B (en) * | 2020-11-26 | 2022-11-18 | 平安普惠企业管理有限公司 | Demand question reply method and device, terminal equipment and storage medium |
| CN112818225A (en) * | 2021-01-27 | 2021-05-18 | 上海明略人工智能(集团)有限公司 | Display method and device of pushed data |
| CN113010654A (en) * | 2021-03-17 | 2021-06-22 | 北京十一贝科技有限公司 | Question reply method and device applied to insurance industry, electronic equipment and medium |
| CN112950233A (en) * | 2021-03-26 | 2021-06-11 | 广东好太太智能家居有限公司 | User message processing method, system, electronic equipment and storage medium |
| CN113343041B (en) * | 2021-06-21 | 2022-05-20 | 北京邮电大学 | Message reply relation judgment system based on graph model representation learning |
| CN113420137A (en) * | 2021-06-29 | 2021-09-21 | 山东新一代信息产业技术研究院有限公司 | Method, device and medium for implementing intelligent question-answering system based on end-to-end framework |
| CN113626568A (en) * | 2021-07-30 | 2021-11-09 | 平安普惠企业管理有限公司 | Man-machine conversation control method and device for robot, computer equipment and medium |
| CN113806508A (en) * | 2021-09-17 | 2021-12-17 | 平安普惠企业管理有限公司 | Multi-turn dialogue method and device based on artificial intelligence and storage medium |
| CN115276697A (en) * | 2022-07-22 | 2022-11-01 | 交通运输部规划研究院 | Coast radio station communication system integrated with intelligent voice |
| CN117114695B (en) * | 2023-10-19 | 2024-01-26 | 本溪钢铁（集团）信息自动化有限责任公司 | Interaction method and device based on intelligent customer service in steel industry |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107766320A (en) * | 2016-08-23 | 2018-03-06 | 中兴通讯股份有限公司 | A kind of Chinese pronoun resolution method for establishing model and device |
| CN109947912A (en) * | 2019-01-25 | 2019-06-28 | 四川大学 | A kind of model method based on paragraph internal reasoning and combined problem answer matches |
| CN109977213A (en) * | 2019-03-29 | 2019-07-05 | 南京邮电大学 | A kind of optimal answer selection method towards intelligent Answer System |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160350653A1 (en) * | 2015-06-01 | 2016-12-01 | Salesforce.Com, Inc. | Dynamic Memory Network |
| US11113598B2 (en) * | 2015-06-01 | 2021-09-07 | Salesforce.Com, Inc. | Dynamic memory network |
-
2019
- 2019-10-25 CN CN201911022722.6A patent/CN111046132B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107766320A (en) * | 2016-08-23 | 2018-03-06 | 中兴通讯股份有限公司 | A kind of Chinese pronoun resolution method for establishing model and device |
| CN109947912A (en) * | 2019-01-25 | 2019-06-28 | 四川大学 | A kind of model method based on paragraph internal reasoning and combined problem answer matches |
| CN109977213A (en) * | 2019-03-29 | 2019-07-05 | 南京邮电大学 | A kind of optimal answer selection method towards intelligent Answer System |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111046132A (en) | 2020-04-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111046132B (en) | Customer service question-answering processing method and system for searching multiple rounds of conversations | |
| CN115238101B (en) | Multi-engine intelligent question-answering system oriented to multi-type knowledge base | |
| CN110321419B (en) | Question-answer matching method integrating depth representation and interaction model | |
| Perez et al. | Dialog state tracking, a machine reading approach using memory network | |
| CN111081220B (en) | Vehicle-mounted voice interaction method, full-duplex dialogue system, server and storage medium | |
| CN110765277B (en) | Knowledge-graph-based mobile terminal online equipment fault diagnosis method | |
| CN111259851B (en) | Multi-mode event detection method and device | |
| CN111353013A (en) | Method and system for realizing intelligent delivery and reception | |
| CN114153955B (en) | Construction method of multi-skill task type dialogue system fusing chatting and common knowledge | |
| CN112287090A (en) | Financial question asking back method and system based on knowledge graph | |
| CN115310551A (en) | Text analysis model training method and device, electronic equipment and storage medium | |
| CN110597968A (en) | Reply selection method and device | |
| CN115982338A (en) | Query path ordering-based domain knowledge graph question-answering method and system | |
| CN114491023A (en) | Text processing method and device, electronic equipment and storage medium | |
| Sathyendra et al. | Helping users understand privacy notices with automated query answering functionality: An exploratory study | |
| CN113420136A (en) | Dialogue method, system, electronic equipment, storage medium and program product | |
| CN115905187B (en) | Intelligent proposition system oriented to cloud computing engineering technician authentication | |
| CN116976341A (en) | Entity identification method, entity identification device, electronic equipment, storage medium and program product | |
| CN116258204A (en) | Industrial safety production violation punishment management method and system based on knowledge graph | |
| CN115730058A (en) | Reasoning question-answering method based on knowledge fusion | |
| CN113177164B (en) | Multi-platform collaborative new media content monitoring and management system based on big data | |
| CN114942981A (en) | Question-answer query method and device, electronic equipment and computer readable storage medium | |
| CN114860908A (en) | Task-based dialogue state tracking method fusing slot association and semantic association | |
| CN112989001A (en) | Question and answer processing method, device, medium and electronic equipment | |
| Card et al. | Automated coding of open-ended survey responses |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |