CN110727778A - Intelligent question-answering system for tax affairs - Google Patents

Intelligent question-answering system for tax affairs Download PDF

Info

Publication number
CN110727778A
CN110727778A CN201910975835.1A CN201910975835A CN110727778A CN 110727778 A CN110727778 A CN 110727778A CN 201910975835 A CN201910975835 A CN 201910975835A CN 110727778 A CN110727778 A CN 110727778A
Authority
CN
China
Prior art keywords
module
user
word
question
model
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.)
Pending
Application number
CN201910975835.1A
Other languages
Chinese (zh)
Inventor
王辉
王宇飞
贾青山
宁博
张益嘉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian Zhonghe Technology Co Ltd
Original Assignee
Dalian Zhonghe Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Dalian Zhonghe Technology Co Ltd filed Critical Dalian Zhonghe Technology Co Ltd
Priority to CN201910975835.1A priority Critical patent/CN110727778A/en
Publication of CN110727778A publication Critical patent/CN110727778A/en
Pending legal-status Critical Current

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
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/10—Tax strategies

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Development Economics (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • Technology Law (AREA)
  • General Business, Economics & Management (AREA)
  • Human Computer Interaction (AREA)
  • Economics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides a tax-oriented intelligent question-answering system, belongs to the field of natural language processing, and aims to solve the problems in the aspect of inquiring tax by a taxpayer in the prior art. The hardware equipment comprises a large server, a mobile terminal device and a computer. The software application program comprises a background maintenance module and a foreground customer service module, wherein the background maintenance module is a database module; the foreground customer service module comprises an intelligent robot module and an artificial customer service module. The invention comprehensively applies the technologies of text classification, similarity calculation and the like in natural language processing, combines the latest attention mechanism method, and can understand the user problems and more accurately push the answers desired by the user. Meanwhile, the method can provide the consultation service for the user all the time in order to meet the actual requirement, and is a main mode of future consultation service.

Description

Intelligent question-answering system for tax affairs
Technical Field
The invention relates to the technical field of natural language processing, in particular to an intelligent question-answering system facing tax affairs.
Background
12366 tax service hotline is the most prominent way for taxpayers to consult tax questions. However, because the taxpayers are numerous and the problems are various and complicated, the answering efficiency of the customer service staff is not high. And because local new policies, preferential policies and the like are changed frequently, the taxpayers cannot know the policies in time, the working pressure of customer service staff is increased, and the service efficiency is reduced.
Disclosure of Invention
According to the technical problems, an intelligent question-answering system facing tax affairs is provided. The invention mainly takes natural language processing technology as the basis, combines the latest knowledge in the directions of statistics, deep learning and the like, and can understand the requirements of users. The invention is provided with the knowledge base with high completeness and strong timeliness, thereby ensuring that the question is answered absolutely and avoiding the question answering. The problem answering service can be uninterruptedly provided for the taxpayers, so that the pressure of customer service staff is effectively relieved, and the consultation efficiency of the taxpayers is improved. Meanwhile, the invention can meet the requirements of deep reform and improvement of service quality of governments, and is a main mode of future consultation service.
The technical means adopted by the invention are as follows:
a tax-oriented intelligent question-answering system, comprising:
the large server is used for storing the knowledge base, the user data and the customer service data and processing messages sent in the using process of the user;
the user terminal equipment is provided with an Android or iOS operating system and is used for acquiring the voice or text message of a client and transmitting the voice or text message to the server for calculation; the server is also used for pushing the related data calculated by the server to the user for the user to select;
the computer is used for the direct communication between the customer service personnel and the user through a conversation interface;
the large server and the computer are both provided with a service software system, and the service software system comprises a background maintenance module and a foreground customer service module;
the user terminal equipment is provided with an application software program which comprises a language algorithm module in the intelligent robot module; meanwhile, the application program of version 6.5.1 of WeChat and above is also installed.
Further, the foreground customer service module comprises an intelligent robot module and an artificial customer service module;
the intelligent robot module comprises a language algorithm module and a problem pushing module;
the language algorithm module further comprises:
the voice conversion module is used for converting the voice information recorded by the user into text information;
the word segmentation module is used for segmenting the text information into word group sequences, and a tool used by the word segmentation module is an LTP word segmentation tool;
the stop word filter is used for removing words which are irrelevant to the practice in the sentences asked by the user and the sentences in the knowledge base according to the existing large-scale stop word list, and the used tool is a HanLP word segmentation tool;
the synonym conversion module is used for carrying out normalization processing on the phrases with the same meaning in practice;
the syntactic analysis module is used for carrying out syntactic analysis on the sentences from which the stop words are removed, and deleting or neglecting words which have little influence on the actual meaning;
its problem propelling movement module still includes:
the similarity calculation module is used for calculating the relation among the keywords; the module integrally uses an ESIM model and introduces a latest attention mechanism method;
the pushing module is used for feeding back the selected problems to the user;
the multi-turn question and answer module is used for acquiring first input information by a background when a user inputs a question, and identifying the acquired first input information so as to determine a preset question; if the information of the first question asked by the user is fuzzy, the background can narrow the range of the questions according to the supplement of the questions during the second question asking and give accurate answers;
the user portrait module is used for depicting the identity image of the same user;
the manual customer service module comprises a communication module used for solving very complicated problems for users on line by manual customer service, and an interface for connecting a customer service end and a user end is also established by the communication module.
Furthermore, the background maintenance module is mainly a database module and comprises operations of adding, deleting, modifying and searching the existing knowledge base, operations of adding, deleting, modifying and searching the chatting records of the manual customer service or the intelligent robot customer service, and marks on the satisfaction degree of the user;
the database module is used for selecting the problems with high frequency, moderate range and high generalization degree as hot problems to be pushed to the user terminal in an enterprise number message mode according to query operation of a large number of users recorded by the database, and is also used for autonomously eliminating invalid problems and adding new policy direction problems, and the pre-training model adopts an ELMo model.
Further, the system needs to perform relevant preamble operations before running: the database module needs to check the existing knowledge items of the knowledge base to ensure moderate tag granularity; training the existing entries of the knowledge base into a pre-training model by using an ELMo model, wherein the language model not only considers the position and frequency of words, but also considers the context relationship; the synonym conversion module checks whether the existing synonym table is complete; the stop word filter checks whether the stop word list is complete.
Furthermore, the ESIM model and the ELMo model both use a Bi-LSTM (bidirectional long-and-short memory model) neural network, the neural network is formed by connecting two layers of forward LSTM networks and backward LSTM networks in series in different directions, and each memory unit also comprises a memory gate, a forgetting gate and an output gate which can be added or deleted.
Further, the memory cell updating process of the Bi-LSTM neural network is as follows:
(1) calculating a forgetting gate: hidden state h input as previous momentt-1And the input word X at the current timetThe output is the value f of the forgetting gatetThen, the calculation formula of the forgetting door is as follows:
ft＝σ(Wf·[ht-1，xt]+bf)
(2) calculating a memory gate: hidden state h input as previous momentt-1And the input word X at the current timetThe output is the value i of the memory gatetAnd temporary cell status
Figure RE-GDA0002281696630000031
The calculation formulas of the memory gate and the temporary cell state are respectively:
it＝σ(Wi·[ht-1，xt]+bi)
Figure RE-GDA0002281696630000032
(3) calculating the unit state at the current moment: input as a value i of a memory gatetForgetting the value f of the doortTemporary cell status
Figure RE-GDA0002281696630000033
Last time cell state Ct-1Output as the cell state C at the current timetThen, the calculation formula of the current time unit state is:
Figure RE-GDA0002281696630000034
(4) calculating the hidden layer states of an output gate and the current time: input as previous hidden statet-1And the input word X at the current timetThe output is the current time unit state CtOutputting the gate and the current time hidden stateThe calculation formulas are respectively as follows:
ot＝σ(Wo·[ht-1，xt]+bo)
ht＝ot*tanh(Ct)。
further, based on the memory unit updating process of the Bi-LSTM neural network, a hidden layer state sequence { h) with the same length as the word sequence is obtained through an LSTM layer0，h1，...，hn-1Predict the k-th word from the first k-1 word sequences over the forward LSTM network:
Figure RE-GDA0002281696630000041
the backward LSTM network predicts the k word from the last N-k word sequences:
Figure RE-GDA0002281696630000042
combining the forward LSTM network and the backward LSTM network by using the maximum likelihood estimation of the logarithm function to finally obtain:
further, when calculating the similarity, the ELMo model selects the output of the last layer of Bi-LSTM and integrates it into a vector, whose expression is:
in the formula (I), the compound is shown in the specification,
Figure RE-GDA0002281696630000045
the output of each layer is represented as a graph,
Figure RE-GDA0002281696630000046
representing the vector of the first layer, the whole expression can also be expressed as:Θtaskindicating the specific task for which γ represents the vector size used to control the ELMo model generation, and S represents the use of the Softmax function as the normalization process between layers.
Further, the similarity calculation module integrally utilizes an ESIM model, including:
the input and coding part adopts a model pre-trained by an ELMo model, inputs the input content into a bidirectional LSTM for coding in order to perform feature extraction on the input content, and finally retains the hidden state values which are respectively recorded as
Figure RE-GDA0002281696630000048
Figure RE-GDA00022816966300000410
In the formula, i and j respectively represent different time, and a and b respectively represent a user question sentence and a target matching sentence;
the local inference model part is used for carrying out difference calculation on the characteristic values obtained by the input and coding parts, namely firstly carrying out similarity calculation on words between two sentences to obtain a two-dimensional similarity matrix:
Figure RE-GDA0002281696630000051
in the above formula, αi，αjAn attention mechanism is applied;
the weights of the question sentence and the matching sentence calculated according to the attention weight are set as the weighted values
Figure RE-GDA0002281696630000052
Figure RE-GDA0002281696630000053
Figure RE-GDA0002281696630000054
And performing difference calculation on the coding value of the corresponding sentence obtained by the input and coding part and the weighted coding value of the corresponding sentence obtained by the local inference model part, namely performing subtraction and multiplication operations on the alignment, and finally splicing the coding value, the weighted coding value, the subtraction value of the alignment and the multiplication value of the alignment together to obtain:
Figure RE-GDA0002281696630000055
the reasoning component part is used for sending the obtained coding information into a BilSTM neural network for calculation and integrating local reasoning information ma，mbAnd context relation, and respectively carrying out average pooling operation V on the results obtained by processing the BilSTM neural networkaveAnd maximum pooling operation VmaxAnd splicing the results to obtain:
Figure RE-GDA0002281696630000056
Figure RE-GDA0002281696630000058
Figure RE-GDA0002281696630000059
V＝[Va，ave，Va，max，Vb，ave，Vb，max]
and the prediction part is used for sending the V of the inference component part into a full connection layer for classification, the activation function adopts tanh, the obtained result is sent to a Softmax layer, and the similarity is finally obtained.
Further, the attention mechanism specifically includes:
outputting the Bi-LSTM model
Figure RE-GDA00022816966300000510
As an input to the attention mechanism, set as E, obtained by the following equation:
E×WQ＝Q；E×WK＝K；E×WV＝V
wherein, in order to improve the expression ability, WQ、WK、WVRespectively representing trainable parameter matrixes, which are different during initialization and can be adjusted during training; q is K and V, and the three are converted from word vectors corresponding to words in the sequence;
when the attention mechanism is calculated, the method mainly comprises the following steps:
the method comprises the following steps: calculating the similarity of Q and each K to obtain a weight;
step two: normalizing the weights by using a Softmax function;
step three: weighting and summing the weight and the corresponding key value V to obtain the final Attention (Q, K, V) ═ ΣiαiViAlpha is finally obtainediI.e. attention weight, we get:
CT＝[α1；...；αT]
compared with the prior art, the invention has the following advantages:
1. the intelligent question-answering system provided by the invention can analyze and understand the questions provided by the user, compare the questions with the existing knowledge in the knowledge base, and return a plurality of most similar questions to the user in real time for the user to select autonomously.
2. The intelligent question-answering system provided by the invention has the advantages that the completeness of the knowledge base is high, the knowledge base is integrated by hundreds of thousands of historical problems, and the coverage is wide; meanwhile, the knowledge base can be updated in real time according to the coming of the new policy.
3. The intelligent question-answering system provided by the invention can be used for counting the current hot problems and policies and making a column interpretation aiming at the hot problems and the hot policies. In addition, on the basis of the prior art, more convenient functions such as social security query and the like can be provided.
4. The intelligent question-answering system provided by the invention adopts a brand-new technology and model: the neural network model is a bidirectional long-time and short-time memory model, the language model is ELMo, and understanding of problem semantics is enhanced through an attention mechanism.
5. The invention provides an intelligent question-answering system which is suitable for users. The system has the advantages that multiple rounds of question answering and user portrait functions are added, so that the user experience is improved, and the experience degree of the user when the platform is applied can be improved.
Based on the reason, the invention can be widely popularized in the fields of artificial intelligence and the like.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a system block diagram of the present invention.
FIG. 2 is a flow chart of the system design of the present invention.
FIG. 3 is a diagram of memory cells in the Bi-LSTM neural network model of the system of the present invention.
FIG. 4 is a schematic illustration of the attention mechanism of the system of the present invention.
FIG. 5 is a flow chart of an ESIM model of the system of the present invention.
Detailed Description
It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict. The present invention will be described in detail below with reference to the embodiments with reference to the attached drawings.
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
The relative arrangement of the components and steps, the numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise. Meanwhile, it should be understood that the sizes of the respective portions shown in the drawings are not drawn in an actual proportional relationship for the convenience of description. Techniques, methods, and apparatus known to those of ordinary skill in the relevant art may not be discussed in detail but are intended to be part of the specification where appropriate. Any specific values in all examples shown and discussed herein are to be construed as exemplary only and not as limiting. Thus, other examples of the exemplary embodiments may have different values. It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, further discussion thereof is not required in subsequent figures.
As shown in fig. 1, in order to effectively process and answer questions posed by users, the invention mainly processes user information in an intelligent robot module, and sometimes relates to a manual customer service module.
The invention provides an intelligent question-answering system facing tax affairs, which comprises:
the large server is used for storing the knowledge base, the user data and the customer service data and processing messages sent in the using process of the user;
the user terminal equipment is provided with an Android or iOS operating system and is used for acquiring the voice or text message of a client and transmitting the voice or text message to the server for calculation; the server is also used for pushing the related data calculated by the server to the user for the user to select; the user terminal equipment is provided with an application software program which comprises a language algorithm module in the intelligent robot module; meanwhile, the application program of version 6.5.1 of WeChat and above is also installed.
The computer is used for the direct communication between the customer service personnel and the user through a conversation interface;
the large server and the computer are both provided with a service software system, and the service software system comprises a background maintenance module and a foreground customer service module;
the foreground customer service module comprises an intelligent robot module and an artificial customer service module;
the intelligent robot module comprises a language algorithm module and a problem pushing module;
the language algorithm module further comprises:
the voice conversion module is used for converting the voice information recorded by the user into text information;
the word segmentation module is used for segmenting text information into word group sequences, and a tool used by the word segmentation module is an LTP word segmentation tool developed by Harbin university of industry;
the stop word filter is used for removing words which are irrelevant to the practice in the sentences asked by the user and the sentences in the knowledge base according to the existing large-scale stop word list, and the used tool is a HanLP word segmentation tool;
the synonym conversion module is used for carrying out normalization processing on the phrases with the same meaning in practice; such as: "give" and "pay" are used in most cases to mean pay, but due to the habit of the user and the problem of input method, many synonyms or spoken words appear. Therefore, on the basis of learning and training word vectors by a machine, a synonym conversion module is additionally added, and the recognition accuracy is improved by adding the synonym conversion module through testing and displaying.
And the syntactic analysis module is used for carrying out syntactic analysis on the sentences from which the stop words are removed, carrying out information screening on the basis of deleting the stop words, and deleting or neglecting words which have little influence on the actual meaning, such as adjectives, adverbs and the like.
The question pushing module is used for pushing the question to the user and further comprises:
the similarity calculation module is used for calculating the relation among the keywords; the module integrally uses an ESIM model and introduces a latest attention mechanism method;
the pushing module is used for feeding back the selected problems to the user; the user can select the problem which the user wants to inquire or seek the help of manual customer service to solve the very complicated problem according to the requirement.
The multi-turn question-answering module is mainly used for avoiding the situation that a user can read a large amount of contents sometimes in a 'question-answer mode'. When a user inputs a problem, a background acquires first input information and identifies the acquired first input information so as to determine a preset problem; the preset question refers to a question corresponding to different answers under different combinations of information input by a user, and each condition combination comprises one or more conditions; initiating a first round of question answering for the determined predetermined question; the following treatments were performed separately in each question and answer round: outputting a question of an unknown condition for the determined predetermined question, obtaining a condition for a user's response; judging whether the obtained condition combination has a corresponding answer or not according to the corresponding relation between the preset condition combination and the answer; if yes, outputting an answer corresponding to the acquired combination of the conditions; if not, the next round of question answering is performed. If the information of the first question asked by the user is fuzzy, the background can narrow the range of the questions according to the supplement of the questions during the second question asking and give accurate answers;
the user portrait module is used for depicting the identity image of the same user; if a user frequently queries questions about "tax stamp", the system may selectively push hot questions about "tax stamp" based on user profile techniques, according to which the user may be interested in. The user profile module provided by the invention mainly comprises modeling of user behaviors. The algorithm models through the historical query behavior of the user to predict the query behavior preference of the user. The existing problems in the knowledge base all have corresponding labels with intermediate granularity, and the labels have certain accuracy and certain generalization capability. To prevent old behavior weights from being too high and new behavior weights from being too low, the algorithm takes into account time decay and time decay.
Order decay equation: scorei+1＝α*scorei+C*weight(0＜α＜1)，
The time decay equation: scoret+1＝scoret*β(0＜β＜1)，
Wherein, alpha and beta are respectively a time attenuation factor and a time attenuation factor; c represents whether the label is present, the presence is 1, and the absence is 0.
The manual customer service module comprises a communication module used for 12366 manual customer service to answer very complicated problems for users on line, and the communication module also creates an interface for connecting a customer service end and a user end.
The background maintenance module is mainly a database module and comprises operations of adding, deleting, modifying and searching the existing knowledge base, operations of adding, deleting, modifying and searching the chatting records of the manual customer service or the intelligent robot customer service and marks on the satisfaction degree of the user;
the database module is used for selecting the problems with high frequency, moderate range and high generalization degree as hot problems to be pushed to the user terminal in an enterprise number message mode according to query operation of a large number of users recorded by the database, and is also used for autonomously eliminating invalid problems and adding new policy direction problems, and the pre-training model adopts an ELMo model.
In the invention, before running, the system needs to perform relevant preamble operations: the database module needs to check the existing knowledge items of the knowledge base to ensure moderate tag granularity; training the existing entries of the knowledge base into a pre-training model by using an ELMo model, wherein the language model not only considers the position and frequency of words, but also considers the context relationship; the synonym conversion module checks whether the existing synonym table is complete; the stop word filter checks whether the stop word list is complete.
As shown in fig. 2, a specific process design of the system of the present invention is mainly as follows:
s1, after entering the platform, the user will send out a question first, the message sent out during the question asking can be voice or text, if it is voice message, the voice conversion module will convert the voice content into text and then enter the word segmentation module.
S2, the word segmentation module can segment the complete sentence into words which are consistent with the understanding of people, and screen out keywords (the keywords are manually input by professional practitioners, and the unique keywords are beneficial to improving the accuracy), and then enter the stop word filter.
S3, stop words such as doubtful words and tone words are removed from the stop word filter, the words do not affect the original meaning of the sentences, the similarity between the sentences can be improved, and the operation time of the background program during operation is reduced.
S4, after the stop word filter is used for processing, the syntactic analysis module performs syntactic analysis on the existing segmented sentences, and filters out some terms of the main information of irrelevant sentences, for example, performing syntactic analysis on the online example of LTP website: 'He shou Tom remove coat'
Called- "Ke (Chinese character) -Ke Ji (sic)
He-Zhu Wei-Zhi
Called- "Fang Yu-" Tang mu (Chinese language-)
Called-move Bin relationship-Na (Na)
Remove the relationship between the root and the middle energizer (remove the meridian-food) -Na
Na-move guest relation-
It can be seen that the removal of the relationships in the shapes also has no effect on the original meaning of the sentence, so the word "remove" can be removed.
S5, after the processing of the syntactic analysis module, the synonym conversion module makes the abbreviation of some proper nouns or the inessential of some words caused by the spoken language so as to reduce the error of the text data when participating in the calculation. For example, "cultural cause construction fee" and "cultural cause construction fee" refer to the same contents; and "decal" and "decal," the references are also identical.
S6, obtaining a word sequence with highly condensed information through the language algorithm module, and calculating the similarity of the word sequence and the content of the knowledge base by the similarity calculation module. Before the system is operated, the platform pre-trains a language model by the Bi-LSTM neural network by using the content in the knowledge base. Through the model, the similarity between the text input by the user and the existing text in the knowledge base can be calculated, and the higher the similarity is, the closer the similarity is.
S7, after the step S6, the pushing problem module pushes a plurality of most similar problems selected according to the calculated similarity to the user, and the user can select the most similar problems and check the solving mode of the problems according to the needs. If there is no answer desired in the pushed content, the user may choose to call human customer service for help.
S8, when the user dialog is started in the current round, the multi-round question answering module starts to operate: the questions asked by the user are recorded by the module. When a user uses software, the user often cannot speak all the content at one time, and the content to be expressed may be divided into a plurality of text sections. At this time, the multi-turn question answering module records all the expressed contents according to steps S1-S5, and searches and predicts the direction that the user may want to inquire in a tree manner according to the keywords in the expressed contents. When the user expresses that the content has five or more different keywords (when inputting multiple pieces of information), the platform will push the user whether the user wants to inquire the content in xx direction, and the user is guided to carry out directional, purposeful and efficient inquiry.
The system is also provided with a database module, the database used by the system is formed by integrating hundreds of thousands of historical problems, and the system has the characteristics of high coverage rate, high completeness, strong timeliness and the like. In order to ensure timeliness, invalidation and policy updating problems need to be eliminated. The policy questions in the database are marked as policy-type questions in the knowledge base, the questions are time-efficient, and once a new policy is out of the platform, the answers of the old related questions are replaced by the new answers. When a question is continuously rated a number of times by different users, the question is listed as "observation area", and if there are still different users who rated a lower rating, the question is temporarily abandoned to the recycle bin and is not involved in subsequent activities. Maintenance personnel will clean the recycle bin regularly.
The design of the invention is mainly characterized by a similarity calculation module which is used for calculating the similarity between texts so as to calculate the relation between keywords. The language model pre-trained in the similarity calculation module adopts an ELMo model. The neural network is a Bi-LSTM neural network (bidirectional long-short time memory model neural network), the neural network is formed by connecting two layers of forward LSTM networks and backward LSTM networks in different directions in series, and the model training effect is better than that of the general LSTM. Bi-LSTM and LSTM are the same, and each memory unit also comprises three parts of a memory gate, a forgetting gate and an output gate which can be added or deleted.
The refresh rule of the memory cells of the Bi-LSTM neural network is shown in FIG. 3, first assuming XtFor the input of the memory cell at time t, htThe hidden state of the memory cell at time t. From cell state Ct, temporary cell state
Figure RE-GDA0002281696630000121
Forget door ftMemory door itOutput gate otAnd (4) forming. By forgetting and memorizing new information in the state of the unit, information useful for calculation at the subsequent moment is transmitted, while useless information is discarded and at each time stepWill input hidden layer state ht. Wherein forgetting, memorizing and outputting h passing through last momentt-1And current input XtCalculated forgetting door ftMemory door itOutput gate otTo control.
The specific updating process is as follows:
(1) calculating a forgetting gate: hidden state h input as previous momentt-1And the input word X at the current timetThe output is the value f of the forgetting gatetThen, the calculation formula of the forgetting door is as follows:
ft＝σ(Wf·[ht-1，xt]+bf)
(2) calculating a memory gate: hidden state h input as previous momentt-1And the input word X at the current timetThe output is the value i of the memory gatetAnd temporary cell status
Figure RE-GDA0002281696630000122
The calculation formulas of the memory gate and the temporary cell state are respectively:
it＝σ(Wi·[ht-1，xt]+bi)
(3) calculating the unit state at the current moment: input as a value i of a memory gatetForgetting the value f of the doortTemporary cell status
Figure RE-GDA0002281696630000124
Last time cell state Ct-1Output as the cell state C at the current timetThen, the calculation formula of the current time unit state is:
Figure RE-GDA0002281696630000125
(4) calculating the hidden layer states of an output gate and the current time: input is hidden at the previous momentLayer state ht-1And the input word X at the current timetThe output is the current time unit state CtThen, the calculation formulas of the output gate and the hidden layer state at the current moment are respectively as follows:
ot＝σ(Wo·[ht-1，xt]+bo)
ht＝ot*tanh(Ct)。
through the memory unit updating process of the Bi-LSTM neural network, the Bi-LSTM neural network realizes the capability of storing, reading and updating the network, simultaneously solves the problem of gradient disappearance, and finally obtains a hidden layer state sequence { h) with the same length as the word sequence through an LSTM layer0，h1，...，hn-1Predict the k-th word from the first k-1 word sequences over the forward LSTM network:
Figure RE-GDA0002281696630000133
the backward LSTM network predicts the k word from the last N-k word sequences:
Figure RE-GDA0002281696630000134
combining the forward LSTM network and the backward LSTM network by using the maximum likelihood estimation of the logarithm function to finally obtain:
Figure RE-GDA0002281696630000131
when the ELMo model is used for calculating the similarity, the output of the last layer of the Bi-LSTM neural network is selected and integrated into a vector, and the expression is as follows:
Figure RE-GDA0002281696630000132
in the formula (I), the compound is shown in the specification,
Figure RE-GDA0002281696630000135
the output of each layer is represented as a graph,
Figure RE-GDA0002281696630000136
representing the vector of the first layer, the whole expression can also be expressed as:
Figure RE-GDA0002281696630000137
Θtaskthe specific task is shown, gamma represents the vector size for controlling the ELMo model to generate, and related research shows that when only the last layer of the ELMo model is taken, the coefficient is beneficial to the optimization of the model, and the accuracy can be greatly improved by adding the coefficient; s denotes that the Softmax function is used as normalization processing between layers. The invention takes only the last layer of the ELMo model to output. The reason is two, specifically as follows:
first, the effect of taking the output of all layers is not significantly improved compared to the effect of taking only the last layer, which also leads to problems such as slow system operation and difficult platform maintenance.
Next, it is generally considered that the output of the lower layer includes syntax information, and the output of the higher layer includes semantic information.
The purpose of introducing the model is to solve semantic information such as ambiguity, so that the invention only takes the last layer of the model to output, thereby not only improving the running speed, but also reducing the time and the storage space required by preprocessing the language model.
In a preferred embodiment of the present invention, the similarity calculation module entirely uses an ESIM model (natural language inference model of an enhanced long-term memory model neural network), and includes:
the input and coding part adopts a model pre-trained by an ELMo model, inputs the input content into a bidirectional LSTM for coding in order to perform feature extraction on the input content, and finally retains the hidden state values which are respectively recorded as
Figure RE-GDA0002281696630000149
Figure RE-GDA00022816966300001410
Figure RE-GDA00022816966300001411
In the formula, i and j respectively represent different time, and a and b respectively represent a user question sentence and a target matching sentence;
the local inference model part is used for carrying out difference calculation on the characteristic values obtained by the input and coding parts, namely firstly carrying out similarity calculation on words between two sentences to obtain a two-dimensional similarity matrix:
Figure RE-GDA00022816966300001412
in the above formula, αi，αjAn attention mechanism is applied;
the weights of the question sentence and the matching sentence calculated according to the attention weight are set as the weighted values
Figure RE-GDA0002281696630000141
Figure RE-GDA0002281696630000142
Figure RE-GDA0002281696630000143
And performing difference calculation on the coding value of the corresponding sentence obtained by the input and coding part and the weighted coding value of the corresponding sentence obtained by the local inference model part, namely performing subtraction and multiplication operations on the alignment, and finally splicing the coding value, the weighted coding value, the subtraction value of the alignment and the multiplication value of the alignment together to obtain:
Figure RE-GDA0002281696630000144
the reasoning component part is used for sending the obtained coding information into a BilSTM neural network for calculation and integrating local reasoning informationMessage ma，mbAnd context relation, and respectively carrying out average pooling operation V on the results obtained by processing the BilSTM neural networkaveAnd maximum pooling operation VmaxAnd splicing the results to obtain:
Figure RE-GDA0002281696630000145
Figure RE-GDA0002281696630000146
Figure RE-GDA0002281696630000147
Figure RE-GDA0002281696630000148
V＝[Va，ave，Va，max，Vb，ave，Vb，max]
and the prediction part is used for sending the V of the inference component part into a full connection layer for classification, the activation function adopts tanh, the obtained result is sent to a Softmax layer, and the similarity is finally obtained.
The main content of Attention mechanism, as shown in FIG. 4, the Attention mechanism in the present invention is essentially referred to as Self-Attention. Its nature can be described as a mapping of a query (Q) to a series of key-value pairs (K, V). Since Self-Attention is paid to the internal structure, Q-K-V is generally considered to be consistent, that is, the contents are all converted from the word vector corresponding to the word in the sequence.
Outputting the Bi-LSTM model
Figure RE-GDA0002281696630000152
As an input to the attention mechanism, set as E, obtained by the following equation:
E×WQ＝Q；E×WK＝K；E×WV＝V
wherein, in order to raise the watchAbility to reach, WQ、WK、WVRespectively representing trainable parameter matrixes, which are different during initialization and can be adjusted during training;
when the attention mechanism is calculated, the method mainly comprises the following steps:
the method comprises the following steps: and calculating the similarity of Q and each K to obtain a weight, and generally selecting operations such as dot product, splicing and the like for calculation. Dot products are selected in the present invention. f (Q, K) ═ QTK
Step two: normalizing the weights by using a Softmax function;
Figure RE-RE-GDA0002281696630000151
step three: weighting and summing the weight and the corresponding key value V to obtain the final Attention (Q, K, V) ═ ΣiαiViAlpha is finally obtainediI.e. attention weight, we get:
CT＝[α1；...；αT]
as a preferred embodiment of the present invention, the programming environment of the present invention is as follows:
CPU：Intel Core I7 8700K 3.70GHz；
memory: custodon 16G;
hard disk: a mechanical hard disk 2T;
a display card: a rainbow Nvidia GeForce GTX 1080Ti 11G;
operating the system: windows 1064 bit;
and (3) developing environment: idea 2019.3.2(jdk1.8), Anaconda (python 3.6);
a database: oracle 11 g;
data format: the formats of csv, txt and bin are main;
programming language: the deep learning part is mainly Python, and the other parts are mainly Java.
Deployment environment main parameters:
CPU：Intel Core I5 3.2GHz；
memory: 16G;
hard disk: 1T;
operating the system: win 764 bits;
a database: oracle 11G;
and (3) developing environment: jdk1.8, Tomcat 8.0;
the maximum number of returned candidate answers to questions is set to 5. The average correct rate of the answer was 82%.
Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. An intelligent question-answering system for tax, which is characterized by comprising:
the large server is used for storing the knowledge base, the user data and the customer service data and processing messages sent in the using process of the user;
the user terminal equipment is provided with an Android or iOS operating system and is used for acquiring the voice or text message of a client and transmitting the voice or text message to the server for calculation; the server is also used for pushing the related data calculated by the server to the user for the user to select;
the computer is used for the direct communication between the customer service personnel and the user through a conversation interface;
the large server and the computer are both provided with a service software system, and the service software system comprises a background maintenance module and a foreground customer service module;
the user terminal equipment is provided with an application software program which comprises a language algorithm module in the intelligent robot module; meanwhile, the application program of version 6.5.1 of WeChat and above is also installed.
2. The tax-oriented intelligent question and answer system according to claim 1, wherein the foreground customer service module comprises an intelligent robot module and an artificial customer service module;
the intelligent robot module comprises a language algorithm module and a problem pushing module;
the language algorithm module further comprises:
the voice conversion module is used for converting the voice information recorded by the user into text information;
the word segmentation module is used for segmenting the text information into word group sequences, and a tool used by the word segmentation module is an LTP word segmentation tool;
the stop word filter is used for removing words which are irrelevant to the practice in the sentences asked by the user and the sentences in the knowledge base according to the existing large-scale stop word list, and the used tool is a HanLP word segmentation tool;
the synonym conversion module is used for carrying out normalization processing on the phrases with the same meaning in practice;
the syntactic analysis module is used for carrying out syntactic analysis on the sentences from which the stop words are removed, and deleting or neglecting words which have little influence on the actual meaning;
its problem propelling movement module still includes:
the similarity calculation module is used for calculating the relation among the keywords; the module integrally uses an ESIM model and introduces a latest attention mechanism method;
the pushing module is used for feeding back the selected problems to the user;
the multi-turn question and answer module is used for acquiring first input information by a background when a user inputs a question, and identifying the acquired first input information so as to determine a preset question; if the information of the first question asked by the user is fuzzy, the background can narrow the range of the questions according to the supplement of the questions during the second question asking and give accurate answers;
the user portrait module is used for depicting the identity image of the same user;
the manual customer service module comprises a communication module used for solving very complicated problems for users on line by manual customer service, and an interface for connecting a customer service end and a user end is also established by the communication module.
3. The tax-oriented intelligent question-answering system according to claim 1 or 2, wherein the background maintenance module is mainly a database module, which comprises operations of adding, deleting, modifying and searching existing knowledge bases, operations of adding, deleting, modifying and searching chatting records of human customer service or intelligent robot customer service, and labeling of user satisfaction;
the database module is used for selecting the problems with high frequency, moderate range and high generalization degree as hot problems to be pushed to the user terminal in an enterprise number message mode according to query operation of a large number of users recorded by the database, and is also used for autonomously eliminating invalid problems and adding new policy direction problems, and the pre-training model adopts an ELMo model.
4. A tax-oriented intelligent question-answering system according to any one of claims 1 to 3, wherein the system requires the relevant preceding operations before running: the database module needs to check the existing knowledge items of the knowledge base to ensure moderate tag granularity; training the existing entries of the knowledge base into a pre-training model by using an ELMo model, wherein the language model not only considers the position and frequency of words, but also considers the context relationship; the synonym conversion module checks whether the existing synonym table is complete; the stop word filter checks whether the stop word list is complete.
5. A tax-oriented intelligent question-answering system according to claim 1, wherein the ESIM model and the ELMo model both use Bi-LSTM (bidirectional long-short term memory model) neural networks, each of the neural networks is composed of two layers of forward LSTM networks and backward LSTM networks in different directions, and each of the memory units also includes three components for adding or deleting a memory gate, a forgetting gate and an output gate.
6. The tax-oriented intelligent question-answering system according to claim 5, wherein the memory unit updating process of the Bi-LSTM neural network is as follows:
(1) calculating a forgetting gate: hidden state h input as previous momentt-1And the input word x at the current timetThe output is the value f of the forgetting gatetThen, the calculation formula of the forgetting door is as follows:
ft＝σ(Wf·[ht-1，xt]+bf)
(2) calculating a memory gate: hidden state h input as previous momentt-1And the input word x at the current timetThe output is the value i of the memory gatetAnd temporary cell status
Figure FDA0002233586500000031
The calculation formulas of the memory gate and the temporary cell state are respectively:
it＝σ(Wi·[ht-1，xt]+bi)
Figure FDA0002233586500000032
(3) calculating the unit state at the current moment: input as a value i of a memory gatetForgetting the value f of the doortTemporary cell statusLast time cell state Ct-1Output as the cell state C at the current timetThen, the calculation formula of the current time unit state is:
Figure FDA0002233586500000034
(4) calculating the hidden layer states of an output gate and the current time: input as previous hidden statet-1And the input word x at the current timetThe output is the current time unit state CtThen the calculation formula of the output gate and the hidden layer state at the current time is divided intoRespectively, the following steps:
ot＝σ(Wo·[ht-1，xt]+bo)
ht＝ot*tanh(Ct)。
7. a tax-oriented intelligent question-answering system according to claim 6, wherein the memory unit updating process based on the Bi-LSTM neural network obtains the hidden state sequence { h) with the same length as the word sequence through an LSTM layer0，h1，...，hn-1Predict the k-th word from the first k-1 word sequences over the forward LSTM network:
Figure FDA0002233586500000035
the backward LSTM network predicts the k word from the last N-k word sequences:combining the forward LSTM network and the backward LSTM network by using the maximum likelihood estimation of the logarithm function to finally obtain:
Figure FDA0002233586500000037
8. a tax-oriented intelligent question-answering system according to claim 6, wherein the ELMo model selects and integrates the output of the last layer of the Bi-LSTM neural network into a vector when calculating the similarity, and the expression is:
Figure FDA0002233586500000038
in the formula (I), the compound is shown in the specification,
Figure FDA0002233586500000039
the output of each layer is represented as a graph,
Figure FDA00022335865000000310
representing the vector of the first layer, the whole expression can also be expressed as:
Figure FDA0002233586500000041
Θtaskindicating the specific task for which γ represents the vector size used to control the ELMo model generation, and S represents the use of the Softmax function as the normalization process between layers.
9. The tax-oriented intelligent question-answering system according to claim 1, wherein the similarity calculation module utilizes an ESIM model in its entirety, comprising:
the input and coding part adopts a model pre-trained by an ELMo model, inputs the input content into a bidirectional LSTM for coding in order to perform feature extraction on the input content, and finally retains the hidden state values which are respectively recorded as
Figure FDA0002233586500000042
Figure FDA0002233586500000043
In the formula, i and j respectively represent different time, and a and b respectively represent a user question sentence and a target matching sentence;
the local inference model part is used for carrying out difference calculation on the characteristic values obtained by the input and coding parts, namely firstly carrying out similarity calculation on words between two sentences to obtain a two-dimensional similarity matrix:
Figure FDA0002233586500000045
in the above formula, αi，αjAn attention mechanism is applied;
the weights of the question sentence and the matching sentence calculated according to the attention weight are set as the weighted values
Figure FDA0002233586500000046
Figure FDA0002233586500000047
Figure FDA0002233586500000048
And performing difference calculation on the coding value of the corresponding sentence obtained by the input and coding part and the weighted coding value of the corresponding sentence obtained by the local inference model part, namely performing subtraction and multiplication operations on the alignment, and finally splicing the coding value, the weighted coding value, the subtraction value of the alignment and the multiplication value of the alignment together to obtain:
Figure FDA0002233586500000049
the reasoning component part is used for sending the obtained coding information into a BilSTM neural network for calculation and integrating local reasoning information ma，mbAnd context relation, and respectively carrying out average pooling operation V on the results obtained by processing the BilSTM neural networkaveAnd maximum pooling operation VmaxAnd splicing the results to obtain:
Figure FDA00022335865000000410
Figure FDA0002233586500000051
Figure FDA0002233586500000052
Figure FDA0002233586500000053
V＝[Va，ave，Va，max，Vb，ave，Vb，max]
and the prediction part is used for sending the V of the inference component part into a full connection layer for classification, the activation function adopts tanh, the obtained result is sent to a Softmax layer, and the similarity is finally obtained.
10. The tax-oriented intelligent question and answer system according to claim 9, wherein the attention mechanism is specifically:
outputting the Bi-LSTM model
Figure FDA0002233586500000054
As an input to the attention mechanism, set as E, obtained by the following equation:
E×WQ＝Q；E×WK＝K；E×WV＝V
wherein, in order to improve the expression ability, WQ、WK、WVRespectively representing trainable parameter matrixes, which are different during initialization and can be adjusted during training; q is K and V, and the three are converted from word vectors corresponding to words in the sequence;
when the attention mechanism is calculated, the method mainly comprises the following steps:
the method comprises the following steps: calculating the similarity of Q and each K to obtain a weight;
step two: normalizing the weights by using a Softmax function;
step three: weighting and summing the weight and the corresponding key value V to obtain the final Attention (Q, K, V) ═ ΣiαiViAlpha is finally obtainediI.e. attention weight, we get:
CT＝[α1；...；αT]。
CN201910975835.1A 2019-10-15 2019-10-15 Intelligent question-answering system for tax affairs Pending CN110727778A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910975835.1A CN110727778A (en) 2019-10-15 2019-10-15 Intelligent question-answering system for tax affairs

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910975835.1A CN110727778A (en) 2019-10-15 2019-10-15 Intelligent question-answering system for tax affairs

Publications (1)

Publication Number Publication Date
CN110727778A true CN110727778A (en) 2020-01-24

Family

ID=69221228

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910975835.1A Pending CN110727778A (en) 2019-10-15 2019-10-15 Intelligent question-answering system for tax affairs

Country Status (1)

Country Link
CN (1) CN110727778A (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111062220A (en) * 2020-03-13 2020-04-24 成都晓多科技有限公司 End-to-end intention recognition system and method based on memory forgetting device
CN111340657A (en) * 2020-02-28 2020-06-26 重庆百事得大牛机器人有限公司 Legal consultation system based on behavior prediction
CN111368191A (en) * 2020-02-29 2020-07-03 重庆百事得大牛机器人有限公司 User portrait system based on legal consultation interaction process
CN111429204A (en) * 2020-03-10 2020-07-17 携程计算机技术（上海）有限公司 Hotel recommendation method, system, electronic equipment and storage medium
CN111625632A (en) * 2020-04-17 2020-09-04 北京捷通华声科技股份有限公司 Question-answer pair recommendation method, device, equipment and storage medium
CN112685564A (en) * 2020-12-28 2021-04-20 广州博士信息技术研究院有限公司 Intelligent science and technology policy classification and pushing method and system
CN113157885A (en) * 2021-04-13 2021-07-23 华南理工大学 Efficient intelligent question-answering system for knowledge in artificial intelligence field
CN114880454A (en) * 2022-07-07 2022-08-09 江西财经大学 Two-stage retrieval type question-answering method and system oriented to psychological support
CN115174949A (en) * 2022-06-30 2022-10-11 广州汇才创新科技有限公司 Projection-based remote live broadcast interaction method and system
CN117852974A (en) * 2024-03-04 2024-04-09 禾辰纵横信息技术有限公司 Online evaluation score assessment method based on artificial intelligence

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106649561A (en) * 2016-11-10 2017-05-10 复旦大学 Intelligent question-answering system for tax consultation service
CN107578106A (en) * 2017-09-18 2018-01-12 中国科学技术大学 A kind of neutral net natural language inference method for merging semanteme of word knowledge
CN109241258A (en) * 2018-08-23 2019-01-18 江苏索迩软件技术有限公司 A kind of deep learning intelligent Answer System using tax field
CN109325780A (en) * 2018-08-24 2019-02-12 安徽讯飞智能科技有限公司 A kind of exchange method of the intelligent customer service system in E-Governance Oriented field
US20190156222A1 (en) * 2017-11-21 2019-05-23 Maria Emma Artificial intelligence platform with improved conversational ability and personality development
CN110298770A (en) * 2019-06-25 2019-10-01 四川长虹电器股份有限公司 A kind of recipe recommendation system

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106649561A (en) * 2016-11-10 2017-05-10 复旦大学 Intelligent question-answering system for tax consultation service
CN107578106A (en) * 2017-09-18 2018-01-12 中国科学技术大学 A kind of neutral net natural language inference method for merging semanteme of word knowledge
US20190156222A1 (en) * 2017-11-21 2019-05-23 Maria Emma Artificial intelligence platform with improved conversational ability and personality development
CN109241258A (en) * 2018-08-23 2019-01-18 江苏索迩软件技术有限公司 A kind of deep learning intelligent Answer System using tax field
CN109325780A (en) * 2018-08-24 2019-02-12 安徽讯飞智能科技有限公司 A kind of exchange method of the intelligent customer service system in E-Governance Oriented field
CN110298770A (en) * 2019-06-25 2019-10-01 四川长虹电器股份有限公司 A kind of recipe recommendation system

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MATTHEW E. PETERS 等: "Deep contextualized word representations" *
QIAN CHEN 等: "Enhanced LSTM for Natural Language Inference" *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111340657A (en) * 2020-02-28 2020-06-26 重庆百事得大牛机器人有限公司 Legal consultation system based on behavior prediction
CN111368191A (en) * 2020-02-29 2020-07-03 重庆百事得大牛机器人有限公司 User portrait system based on legal consultation interaction process
CN111368191B (en) * 2020-02-29 2021-04-02 重庆百事得大牛机器人有限公司 User portrait system based on legal consultation interaction process
CN111429204A (en) * 2020-03-10 2020-07-17 携程计算机技术（上海）有限公司 Hotel recommendation method, system, electronic equipment and storage medium
CN111062220A (en) * 2020-03-13 2020-04-24 成都晓多科技有限公司 End-to-end intention recognition system and method based on memory forgetting device
CN111625632A (en) * 2020-04-17 2020-09-04 北京捷通华声科技股份有限公司 Question-answer pair recommendation method, device, equipment and storage medium
CN112685564A (en) * 2020-12-28 2021-04-20 广州博士信息技术研究院有限公司 Intelligent science and technology policy classification and pushing method and system
CN113157885A (en) * 2021-04-13 2021-07-23 华南理工大学 Efficient intelligent question-answering system for knowledge in artificial intelligence field
CN115174949A (en) * 2022-06-30 2022-10-11 广州汇才创新科技有限公司 Projection-based remote live broadcast interaction method and system
CN115174949B (en) * 2022-06-30 2024-02-23 广州汇才创新科技有限公司 Remote live broadcast interaction method and system based on projection
CN114880454A (en) * 2022-07-07 2022-08-09 江西财经大学 Two-stage retrieval type question-answering method and system oriented to psychological support
CN117852974A (en) * 2024-03-04 2024-04-09 禾辰纵横信息技术有限公司 Online evaluation score assessment method based on artificial intelligence

Similar Documents

Publication Publication Date Title
CN110727778A (en) Intelligent question-answering system for tax affairs
CN109885672B (en) Question-answering type intelligent retrieval system and method for online education
CN111783474B (en) Comment text viewpoint information processing method and device and storage medium
CN108597519B (en) Call bill classification method, device, server and storage medium
WO2019153737A1 (en) Comment assessing method, device, equipment and storage medium
CN109902159A (en) A kind of intelligent O&M statement similarity matching process based on natural language processing
CN109460457A (en) Text sentence similarity calculating method, intelligent government affairs auxiliary answer system and its working method
CN109829052A (en) A kind of open dialogue method and system based on human-computer interaction
CN109086265B (en) Semantic training method and multi-semantic word disambiguation method in short text
US11210286B2 (en) Facet-based conversational search
CN109710766B (en) Complaint tendency analysis early warning method and device for work order data
CN110083837B (en) Keyword generation method and device
US20220138770A1 (en) Method and apparatus for analyzing sales conversation based on voice recognition
CN109325780A (en) A kind of exchange method of the intelligent customer service system in E-Governance Oriented field
US20230169271A1 (en) System and methods for neural topic modeling using topic attention networks
Su et al. Attention-based response generation using parallel double Q-learning for dialog policy decision in a conversational system
CN113988086A (en) Conversation processing method and device
CN115062003A (en) Cloud ERP community generation type question-answering method based on GPT2
CN117709358A (en) Dialogue response method, device, equipment and medium of insurance intelligent question-answering system
CN114722833A (en) Semantic classification method and device
Aattouri et al. Modeling of an artificial intelligence based enterprise callbot with natural language processing and machine learning algorithms
CN117668199A (en) Intelligent customer service question-answer prediction and recommendation dialogue generation method and device
CN117648915A (en) Question and answer scoring method and system based on knowledge graph
CN110633363B (en) Text entity recommendation method based on NLP and fuzzy multi-criterion decision
Dikshit et al. Automating Questions and Answers of Good and Services Tax system using clustering and embeddings of queries

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20200124