CN112287090A - Financial question asking back method and system based on knowledge graph - Google Patents

Abstract

The invention relates to a financial question asking method and system based on knowledge graph, include carrying on the situation classification of scene to the content that users input according to asking the classification model, carry on the question asking operation when the classification result is the scene situation asking the question when the content is not clear or is important content; constructing a knowledge graph aiming at a financial question-answer knowledge system; matching question statements based on the knowledge graph; obtaining a question-back entity and an entity type structure, combining a sentence pattern template library to obtain a question-back sentence pattern, and generating a question-back sentence; constructing a sound prediction model and an expression prediction model based on a neural network classifier and acquiring sound and visual expressions; the chat robot in the financial field combines the question-back form, the sound and the visual expression to ask the user back. The invention can make the question-answering system fully know the user information and the question information of the user, and provide real and accurate question-answering when the question is ambiguous, thereby making the answer more correct and improving the ability of the question-answering system to solve the user question.

Description

Financial question asking back method and system based on knowledge graph

[ technical field ] A method for producing a semiconductor device

The invention relates to the technical field of artificial intelligence, in particular to a financial question asking method and system based on a knowledge graph.

[ background of the invention ]

In the question-answering system in the financial field, people often input contents to ask questions, and then the question-answering robot answers directly. However, when the questions posed by people are ambiguous or ambiguous, the question-answering robot is easy to answer the questions and cannot solve the real questions of the users. When the user does not know what should be asked or the user does not understand the question deeply, the question is likely to lack part of information, and at this time, the conversation robot is difficult to work effectively. Aiming at the problems, the conversation robot needs to feed back or ask back under proper conditions, so that the conversation robot can better interact with people, and the conversation robot can clearly know the real problems of people, so that the problems are effectively solved. The dialogue robot is favorable for knowing more information and confirming user information in a back questioning mode, and has an important function. In order to enable the conversation robot to effectively solve the problems of people, a question-answering mechanism in a question-answering system is more accurate and feasible, so that the design of more accurately answering question-answering questions in the financial field is an important problem to be solved.

The invention discloses an intelligent question-answering method, an intelligent question-answering device and a computer storage medium based on a knowledge graph in patent CN110334272A, and the intelligent question-answering method based on the knowledge graph comprises the following steps: the method comprises the steps of obtaining a question-answer data set, constructing a knowledge graph relational data set according to the question-answer data set, preprocessing the knowledge graph relational data set to obtain a logic question-answer data set, extracting a logic question data set in the logic question-answer data set, calculating question similarity among data in the logic question data set, establishing a webpage chain interface, preprocessing after receiving a question input by a user, calculating question similarity between the question input by the user and data in the logic question data set, judging the size relation between the question similarity set and a preset question threshold value, and outputting answers of the question input by the user. The invention obtains answers by matching questions according to the question similarity, and can answer questions and answers with clear logical expression. However, when the questions presented by people are ambiguous or ambiguous, the method is easily matched to irrelevant questions and easy to answer questions without solving the real questions of the users, and the method does not relate to the question asking back of the knowledge in the financial field. Therefore, how to feed back or ask questions in a proper situation to guide people to express real problems is very important.

[ summary of the invention ]

The invention provides a financial question-answering method based on a knowledge graph, which is used for enabling a question-answering system to fully know user information and question information of a user and to ask a question when the question is ambiguous, so that the answer is more correct, and the ability of the question-answering system for solving the user question is improved.

The invention provides a financial question asking method based on a knowledge graph, which mainly comprises the following steps:

performing scene situation classification on the content input by the user according to the question classification model, and performing question-back operation when the classification result is the scene situation of question-back when the content is not clear or is important; constructing a knowledge graph aiming at a financial question-answer knowledge system; matching question statements based on the knowledge graph; obtaining a question-back entity and an entity type structure, combining a sentence pattern template library to obtain a question-back sentence pattern, and generating a question-back sentence; establishing a sound prediction model and an expression prediction model based on a neural network classifier to obtain sound and visual expressions; the financial robot asks the user in return in combination with question patterns, sound and visual expressions.

Further optionally, in the method as described above, the classifying the scene condition of the content input by the user mainly includes the following steps:

judging whether the input content of the user has ambiguity or not, including segmenting the input content, judging whether the question of the user can be identified or not through an ambiguity resolution algorithm, further judging the probability of ambiguity or not when the ambiguity resolution algorithm judges that the input content has ambiguity, and when two ambiguous segmentation results are obtained, one is larger than the other and exceeds a certain threshold, the question does not need to be asked, otherwise, the question needs to be asked for the user;

if the content input by the user is not ambiguous, judging whether the granularity of the input content of the user is proper or not, including obtaining the content input by the user, carrying out named entity recognition, obtaining entity words of the content, matching which layer of the words in the synonym forest is obtained through the synonym forest, and if the number of layers is less than or equal to 2, determining that the granularity is not proper and asking the user for question.

Further optionally, the constructing of the knowledge graph mainly comprises the following steps:

constructing a hierarchical knowledge system, wherein the system comprises a set large class, a set small class and elements; extracting keywords of input content based on a keyword automatic extraction technology of sequence labeling; converting the keywords into word vectors based on lookup table embedding of a pre-training model; inputting a word2vec module to obtain similar words or different expressions of each keyword; and judging whether the words co-occur in the same sentence, and extracting the relation between the problem and the subclass or the relation between the subclass and the element when the co-occurrence probability is greater than the preset times.

Further optionally, the performing question-back matching based on the knowledge graph mainly includes:

constructing a question classification model which is a text classification model capable of classifying the scene condition of an input text;

acquiring a training corpus, including acquiring a financial conversation corpus and a scene label obtained after the financial conversation corpus is subjected to scene condition labeling;

training a question classification model according to the training corpus, and repeatedly iterating until convergence to obtain a trained question classification model;

and the trained question classification model classifies the scene condition of the content input by the user and judges whether question-returning operation is needed or not.

Further optionally, the obtaining a question-back entity and an entity type structure further comprises,

collecting financial questioning data of a user as a training set text, collecting a word vector set of entity type labels signed manually, and performing word segmentation pretreatment on the training set text;

obtaining a word vector represented by a distributed form of words in a text of a training set by using a word vector space constructed by a word2vec tool;

and pre-training a BERT model of Google on a BLSTM-CRF model by utilizing the training concentrated word vectors and the existing entity type labels of each word vector to obtain a BERT-BilsTM-CRF-NER entity type prediction model, and analyzing the input content by adopting the model and obtaining the entity and the entity type.

Further optionally, the question-back entity may further obtain similarity between the knowledge graph pair and the embedding vector according to the knowledge graph relationship, obtain a key vocabulary of the user input content according to the knowledge graph relationship, convert the key vocabulary into the embedding vector, perform similarity calculation on a plurality of embedding vectors, and if a word meeting a similarity threshold is found, indicate that the element is mentioned; otherwise, element deletion exists, and the corresponding element key words are taken out to be used as question-back entities.

Further optionally, the obtaining the question-back pattern further includes comparing and matching the obtained entity type structure with the pattern template library to obtain the question-back pattern corresponding to the entity type structure.

Further optionally, the generating a question return further includes, based on the question return entity and the question return pattern, calling a GPT2 model to generate a question return; or generating question-reversing sentences based on question-reversing entities, question-reversing patterns and albert models.

Further optionally, the acquiring the sound and the visual expression further includes designing the sound and the visual expression according to a scene condition of content input by a user, and the method mainly includes:

constructing a sound prediction model based on entity and sound corresponding data by utilizing a neural network classifier; predicting sound using the sound prediction model for an input entity of the user; and constructing an expression prediction model for the entity and the data corresponding to the sound expression by using a neural network classifier, and predicting the expression for the user input entity by using the expression prediction model.

The invention provides a financial question asking system based on knowledge graph, the system includes:

the scene condition classification module is used for classifying the scene conditions of the content input by the user;

the knowledge graph relation module can judge and classify the elements of the content input by the user;

the question return generation module is used for acquiring question return entities, entity type structures and question return patterns and finally generating question return;

and the sound expression design module is used for designing the sound and visual expression of the matched financial robot in the back questioning process according to the scene condition.

The technical scheme provided by the embodiment of the invention has the following beneficial effects:

the method provided by the invention is beneficial for the question-answering system to know more information and confirm the question content of the user, and when the content provided by the user is ambiguous and fuzzy, the question is asked in reverse in time, so that the content which is not required by the user is avoided; in addition, when the user proposes the relevant problems of the important contents, the user is given a question back confirmation, so that the user is ensured to perform relevant important operations, and the possibility of misoperation is reduced. By the method, the question-answering system can answer financial related questions provided by the user more correctly, and the ability of the question-answering system for solving the user problems is improved.

[ description of the drawings ]

FIG. 1 is a schematic flow chart diagram of a method for knowledge-graph based questioning financial questions according to the present invention;

FIG. 2 is a schematic diagram of a financial question asking system based on knowledge graph according to the present invention.

[ detailed description ] embodiments

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in detail with reference to the accompanying drawings and specific embodiments.

FIG. 1 is a flow chart of a financial question asking method based on knowledge graph according to the present invention. As shown in fig. 1, the financial question asking method based on knowledge graph of this embodiment may specifically include the following steps:

step 1, carrying out scene condition classification on the content input by the user according to a question classification model, and when the classification result is ambiguous content or important content, carrying out question reversing operation.

Specifically, firstly, the input content is segmented, whether the question of the user can be identified is judged through an ambiguity resolution algorithm, and when the ambiguity resolution algorithm judges that ambiguity exists in the input content, the probability of the ambiguity is further judged. When two ambiguous word segmentation results are larger than one another by a certain threshold value, a question does not need to be asked, wherein the threshold value can be set by a person skilled in the art according to actual requirements. Otherwise, the user is asked questions. For example, for the input content being a question, "will students organize charitable activities? "includes two different directions of understanding: first, is student | organize charitable activities? I.e. the subject is a student, two of which, the student will organize charitable activities? The subject is a student meeting, and the subject needs to be asked back to confirm the question, namely whether the charitable activity is organized by students or student meetings; some ambiguities may be resolved and resolved by disambiguation algorithms, but some are disambiguated. For ambiguity that cannot be resolved, a question is asked backwards. The ambiguity resolution algorithm analyzes whether two or more kinds of word segmentation are correct or not. And judging whether ambiguity exists. If this is the case, a question is asked. The specific disambiguation algorithm can refer to a combinatorial disambiguation algorithm for Chinese word segmentation.

If the input content of the user does not have ambiguity, further judging whether the granularity of the input content of the user is proper, wherein the granularity is the detail degree of the input content, mainly comprising the steps of obtaining the content input by the user, carrying out named entity recognition, obtaining entity words of the input content, and matching which layer of the words in a synonym forest through the synonym forest. The words of the synonym forest can be divided into 6 layers according to the letter composition of the serial number, and the concept is thinner as the layer number is deeper. If the number of layers is the first and second layers, the description is not detailed enough, and further questions are needed. The situation that the granularity of the input content is too coarse includes, for example, the user says that he lives in China, and the question and answer robot is intended to ask which city and even which street and house number he lives in China, and when the granularity of the input content is not appropriate, the user needs to ask the question again.

And after the fact that the dialogue content of the user needs to be further asked reversely is rapidly and preliminarily judged through the synonym forest. Further, it is determined for which portion of the content the query should be made with respect to whether the content in the user input matches the content in the knowledge-graph triplets. The specific scheme refers to step two.

If the input content of the user is appropriate in granularity, that is, detailed enough, it is determined whether the answer is important content, that is, whether the name of an important event such as money, time, place, and the like exists in the input content of the user is further analyzed. And judging which word needs to be questioned through the matching of the knowledge graph. For example, can a question be loaned by 20.5 ten thousand? The question is designed for large amounts of money, which is important, and requires a second confirmation of the loan amount, such as a question back to determine if the loan amount is 20.5 ten thousand.

When the content input by the user is ambiguous, namely the content is ambiguous, or the granularity of the problem is not appropriate; and when the content input by the user is judged to be important content, the content input by the user needs to be confirmed by asking back.

And step two, constructing a knowledge graph aiming at the financial question-answer knowledge system.

The process further comprises the steps of: constructing a hierarchical knowledge system, namely generally classifying all the problems into which large categories, wherein each category of problems comprises which small categories, and which elements are arranged below small points, and the large categories, the small categories and the elements are set by professionals in the field through manual marking according to business needs; the keyword automatic extraction technology based on sequence labeling can extract the keywords of the problems, the subclasses and the elements from the original corpus to obtain the important keywords of the elements; converting the keywords into word vectors based on lookup table embedding of a pre-training model; inputting a word2vec module to obtain similar words or different expressions of each keyword, for example, the word 'finance' can obtain synonyms 'stock market', 'financing', 'economy' and the like of the word through a word2vec algorithm. And judging whether the words co-occur in the same sentence, and extracting the relation between the problem and the subclass or the relation between the subclass and the element when the co-occurrence probability is greater than the preset times. For example, the 'stock market' is often presented simultaneously with the 'big disc' as the big class, so when the stock market is set as the big class, it can be known that the 'stock market big disc' is the small class, and if the big disc and the big disc are increased in amplitude, and the 'stock market big disc increased amplitude' is considered as the element if the big disc and the big disc are determined to be frequently co-present through word frequency statistics.

Specifically, for the question 'how to handle commercial loan', the question can be matched to be 'loan' through the matching of the knowledge graph, and further, the subclass associated with the matched question can be further matched to be: the loan process, the loan amount and the advance repayment are automatically extracted as the triples of loan, loan process, subclass, loan amount, subclass, loan, advance repayment and subclass; further, accurate answers of each subclass are searched, such as loan amount, which refers to the amount of money, different amounts of money and different interest, and the specific interest is what. Further, it is necessary to know whether the loan is an enterprise loan or a personal loan, the loan purpose, whether there is a mortgage, and the like. At this level, the triplets < loan amount, currency type, element >, < loan amount, amount interest, element > and so on are thus constructed.

And step three, performing question reversing matching based on the knowledge graph.

In the question and answer process, the knowledge graph constructed in the step can be used for judging whether a question needs to be generated or not according to whether the elements of the question are complete or not in the question and answer scene, namely, a question classification model is constructed firstly. The system is used for judging the category of the input content of the user in the knowledge graph; the step can use a text classification model to carry out recognition judgment; after the field of the input content of the user is obtained, whether the elements are complete or not is judged, and judgment can be carried out through keyword retrieval. In the above decisions made from user input, if no feedback of a certain type is made explicitly at any one step, a question back confirmation needs to be initiated to the questioner.

In the above steps, the constructing of the question classification model mainly includes the following steps:

constructing a question classification model which is a text classification model capable of classifying the scene condition of an input text; specifically, the text classification models that can be used in this embodiment include models such as FastText, LSTM, BERT, and the like, and are used for text classification.

The training of the question classification model further comprises the following steps: firstly, acquiring a training corpus, including acquiring a financial conversation corpus and a scene label obtained after scene condition marking is carried out on the financial conversation corpus; the scenes refer to business scenes existing on all business lines in the financial field, such as various scenes of borrowing, investment, stock consultation, fund and the like. In this embodiment, the financial conversation corpus may be from an open financial corpus, or actually collected on various large websites such as forum news, and the collected content should cover various common questions and answers in the financial field, so as to comprehensively cover various scene conditions, and the plate names therein may be used as the automatic labeling content of these corpora. And then, labeling the obtained dialogue linguistic data manually aiming at the missing scenes, and labeling the scene condition corresponding to each dialogue. And manually checking the conversation content, judging the real scene situation according to the conversation content, and marking a scene label for the conversation.

And training an LSTM model or a BERT model according to the training corpus, and obtaining a trained question classification model through repeated iteration until convergence.

And the trained question classification model classifies the scene condition of the content input by the user, judges the field scene to which the user question belongs, matches the question layer in the knowledge map according to the field scene, matches the subclass through the question layer, and finally matches the element layer, and performs question reversing operation on the input of the user when the corresponding answer is not matched.

And step four, obtaining the question-reversing entity and the entity type structure, combining the sentence pattern template library to obtain a question-reversing sentence pattern, and generating the question-reversing sentence.

The question-back entity and entity type structure further includes,

collecting financial questioning data of a user as a training set text, collecting a word vector set of entity type labels signed manually, and performing word segmentation pretreatment on the training set text;

obtaining a word vector represented by a distributed form of words in a text of a training set by using a word vector space constructed by a word2vec tool;

and pre-training a BERT model of Google on a BLSTM-CRF model by utilizing training concentrated word vectors and existing entity type labels of each word vector to obtain the BERT-Bilstm-CRF-NER entity type prediction model, wherein the model is popular entity identification open source software on github and can be downloaded, installed and used. By adopting the model for analysis, corresponding entities and entity types can be obtained.

Specifically, if the user input is "i want to loan 20 ten thousand", the prediction model based on the entity category of the open source tool BERT-BilSTM-CRF-NER is matched to the fact that the entity is a loan, the type is a fund type, and the fund type needs to be confirmed again.

In this embodiment, the question-back entity may further obtain similarity between the knowledge graph pair and the embedding vector according to the knowledge graph relationship, obtain a key vocabulary of the user input content according to the knowledge graph relationship, convert the key vocabulary into the embedding vector, perform similarity calculation on a plurality of embedding vectors, and if a word meeting a similarity threshold is found, indicate that the element is referred to; otherwise, element deletion exists, and the corresponding element key words are taken out to be used as question-back entities.

For example, for the question "can loan 20 ten thousand? ", the question may be judged about the loan by the question classification model; associating subclasses including loan flow, loan amount, advance repayment and the like according to the knowledge graph of the loan, and searching by keywords to obtain the problem related to the loan amount; inquiring the knowledge map, and finding out the elements for confirming the answer of the loan amount question, wherein the elements comprise two items of client identity and loan application; converting the vocabulary of the question into word embedding vectors, and then carrying out similarity calculation with the embedding vectors of the two keywords of client identity and loan purpose.

And the step of obtaining the question-back patterns further comprises the step of comparing and matching the obtained entity type structures with the pattern template library to obtain the question-back patterns corresponding to the entity type structures.

The step of generating the question-reversing sentence further comprises the step of calling a GPT2 model to generate the question-reversing sentence based on the question-reversing entity and the question-reversing pattern; or generating question-reversing sentences based on question-reversing entities, question-reversing patterns and albert models.

Specifically, for the GPT2 model, when the user input content is "20 ten thousand" and the question-asking entity is "20 ten thousand", the question-asking pattern is "[ amount ] + is + [ item ] +? ", the entity and sentence pattern are converted into data in a form recognizable by the GPT2 model and input into the GPT2 model, and the question" 20 ten thousand are the loan amount? ".

The albert model is one of the common models in the current natural language processing field, which should be trained and used before. The question-reversing entity, the question-reversing pattern and the input of the user constitute the input information for generating the question-reversing sentence, and the generation of the question-reversing sentence based on the albert model further comprises the following steps,

data construction, namely, sample data needs to be constructed aiming at the existing problems and question-reversing sentence data pairs; in this embodiment, the category, the sub-question, and the element of the question are confirmed by the classifier according to the input content of the user.

Performing word vector matching on the vocabulary of the question and the element vocabulary; the user of the process confirms which specific element is being asked according to which question-reversing sentence.

Obtaining sentence pattern information through a discrimination model; in this case, samples in the form of (question, element entity, sentence, question-back) can be obtained.

Further, constructing a generative model albert; the basic framework of the model is an albert model proposed in Google 2019, the model is a seq2seq model constructed based on a transformer feature extractor, mass corpora are used in advance to perform pre-training based on two tasks of next sense prediction and mask language model, and model basic weight is generated. In specific application, other network structures can be built on the albert model based on the transfer learning idea, and field data is used for model fine tuning training, so that a good application effect can be obtained for general natural language processing tasks. The generating task in this embodiment conforms to a typical encoder-decoder framework, the encoder part only needs to initialize a multi-head multi-layer transformer of the encoder part by using a pre-trained albert model, for the decoder end, since data crossing needs to be prevented, the decoder end needs to make a mask on the basis of the multi-head multi-layer transformer, then performs module fine tuning training according to a seq2seq model force strategy by using albert pre-training weights as initial weight values, and trains albert by using a sample processed in advance as fine tuning corpus.

The albert model generates a question sentence according to the question sentence text, the element entities and the sentence pattern.

And fifthly, constructing a sound prediction model and an expression prediction model based on the neural network classifier to obtain sound and visual expressions.

Acquiring the sound and the visual expression further comprises designing the sound and the visual expression according to the scene condition of the content input by the user, and the method mainly comprises the following steps:

constructing a sound prediction model based on entity and sound corresponding data by utilizing a neural network classifier; predicting sound using the sound prediction model for an input entity of the user; an expression prediction model is built for data corresponding to an entity and a sound expression by Using a Neural network classifier, an expression is predicted for the user input entity by Using the expression prediction model, the prediction method is realized by adopting a method of a treatise of Dank Learning, Generating means Using Deep Neural, combining with the realized open source project, Generating an expression, and Generating a counter expression in a counter question. The neural network-based classifier can also be analyzed by adopting a hundred-degree speech synthesis AipSpeech interface. The key emphasis effect of the question is achieved by adjusting the relation among the speed of speech, the volume and the tone. For example, when the user is asked for questions, the key words asked for questions are played to the user in a larger volume, and the tone is increased, so that a more distinctive question asking effect and an interactive feeling are achieved.

For example, when the user inputs 20 thousands of loans, the identified scene condition is that the content input by the user is important content and is an important scene of loan application, a reminding sound effect is designed to prompt the loan amount, and 20 thousands of loans are subjected to volume increase and tone increase. A conspicuous expression is set near the loan amount "20 ten thousand" while making the text of "20 ten thousand" conspicuous. Thereby achieving better question-return reminding effect. And finally, the financial robot asks the user in a return mode by combining the return sentence pattern, the sound and the visual expression.

It should be understood by those skilled in the art that each module in the above system may be executed by a program code based on a computer or an intelligent data processing chip, in this embodiment, the system program is finally imported into a robot capable of interacting with a user and running, and in this embodiment, the robot is a financial robot, and the user performs question answering and obtains information required by the user through interaction with the financial robot.

Programs for implementing the information governance of the present invention may be written in computer program code for carrying out operations of the present invention in one or more programming languages, including an object oriented programming language such as Java, python, C + +, or a combination thereof, as well as conventional procedural programming languages, such as the C language or similar programming languages. The program code may execute entirely on the student computer, partly on the student computer, as a stand-alone software package, partly on the student computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the student computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider).

In the embodiments provided in the present invention, it should be understood that the disclosed system, apparatus and method may be implemented in other ways. For example, the above-described device embodiments are merely illustrative, and for example, the division of the units is only one logical functional division, and other divisions may be realized in practice.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, or in a form of hardware plus a software functional unit. The integrated unit implemented in the form of a software functional unit may be stored in a computer readable storage medium. The software functional unit is stored in a storage medium and includes several instructions to enable a computer device (which may be a personal computer, a server, or a network device) or a processor (processor) to execute some steps of the methods according to the embodiments of the present invention.

And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (10)

1. A financial question asking method based on knowledge graph, which is characterized in that the method comprises the following steps:

performing scene situation classification on the content input by the user according to the question classification model, and performing question-back operation when the classification result is the scene situation of question-back when the content is not clear or is important; constructing a knowledge graph aiming at a financial question-answer knowledge system; matching question statements based on the knowledge graph; obtaining a question-back entity and an entity type structure, combining a sentence pattern template library to obtain a question-back sentence pattern, and generating a question-back sentence; and synthesizing the question-reversing sentences into the speech and display effect of the question-reversing related types, and performing question-reversing on the user.

2. The method of claim 1, wherein said scene-wise classifying the user-entered content comprises essentially of the steps of:

judging whether the input content of the user has ambiguity or not, including segmenting the input content, judging whether the question of the user can be identified or not through an ambiguity resolution algorithm, further judging the probability of ambiguity or not when the ambiguity resolution algorithm judges that the input content has ambiguity, and when two ambiguous segmentation results are obtained, one is larger than the other and exceeds a certain threshold, the question does not need to be asked, otherwise, the question needs to be asked for the user;

if the content input by the user is not ambiguous, judging whether the granularity of the input content of the user is proper or not, including obtaining the content input by the user, carrying out named entity recognition, obtaining entity words of the content, matching which layer of the words in the synonym forest is obtained through the synonym forest, and if the number of layers is less than or equal to 2, determining that the granularity is not proper and asking the user for question.

3. The method of claim 2, wherein the constructing a knowledge-graph essentially comprises the steps of:

constructing a hierarchical knowledge system, wherein the system comprises a set large class, a set small class and elements; extracting keywords of input content based on a keyword automatic extraction technology of sequence labeling; converting the keywords into word vectors based on lookup table embedding of a pre-training model; inputting a word2vec module to obtain similar words or different expressions of each keyword; and judging whether the words co-occur in the same sentence, and extracting the relation between the problem and the subclass or the relation between the subclass and the element when the co-occurrence probability is greater than the preset times.

4. The method of claim 3, wherein the knowledge-graph based question-back matching essentially comprises:

constructing a question classification model which is a text classification model capable of classifying the scene condition of an input text;

acquiring a training corpus, including acquiring a financial conversation corpus and a scene label obtained after the financial conversation corpus is subjected to scene condition labeling;

training a question classification model according to the training corpus, and repeatedly iterating until convergence to obtain a trained question classification model;

and the trained question classification model classifies the scene condition of the content input by the user and judges whether question-returning operation is needed or not.

5. The method of claim 4, wherein said obtaining a challenge-back entity and entity type structure further comprises,

collecting financial questioning data of a user as a training set text, collecting a word vector set of entity type labels signed manually, and performing word segmentation pretreatment on the training set text;

obtaining a word vector represented by a distributed form of words in a text of a training set by using a word vector space constructed by a word2vec tool;

and pre-training a BERT model of Google on a BLSTM-CRF model by utilizing the training concentrated word vectors and the existing entity type labels of each word vector to obtain a BERT-BilsTM-CRF-NER entity type prediction model, and analyzing the input content by adopting the model and obtaining the entity and the entity type.

6. The method according to claim 4 or 5, wherein the question-backing entity can also obtain similarity between the knowledge-graph relationship pair and the embedding vector, obtain key words of user input contents according to the knowledge-graph relationship, convert the key words into the embedding vector, perform similarity calculation on a plurality of embedding vectors, and if words meeting a similarity threshold are found, explain elements are mentioned; otherwise, element deletion exists, and the corresponding element key words are taken out to be used as question-back entities.

7. The method of claim 1, wherein the obtaining a question-back pattern further comprises obtaining a question-back pattern corresponding to the entity type structure by comparing and matching the obtained entity type structure with a pattern template library.

8. The method of claim 1, wherein the generating a question return further comprises invoking a GPT2 model for question return generation based on the question return entity and the question return pattern; or generating question-reversing sentences based on question-reversing entities, question-reversing patterns and albert models.

9. The method according to claim 1, wherein the step of synthesizing the question-reversing sentence according to the generated question-reversing sentence into a question-reversing related type of voice and a display effect to question the user in a reverse manner mainly comprises:

the Baidu speech synthesis AipSpeech interface is adopted for analysis, and the effect of emphasizing key question is achieved by adjusting the speed, volume and tone.

10. A financial question asking system based on knowledge-graph, the system comprising:

the scene condition classification module is used for classifying the scene conditions of the content input by the user;

the knowledge graph relation module can judge and classify the elements of the content input by the user;

the question return generation module is used for acquiring question return entities, entity type structures and question return patterns and finally generating question return;

and the voice expression design module is used for designing matched financial robots according to the question-reversing sentences to perform voice and display during question-reversing.

Images