CN107958091A

CN107958091A - A kind of NLP artificial intelligence approaches and interactive system based on financial vertical knowledge mapping

Info

Publication number: CN107958091A
Application number: CN201711459522.8A
Authority: CN
Inventors: 马天平; 侯玥
Original assignee: Beijing Beta Technology Co Ltd
Current assignee: Beijing Beta Technology Co Ltd
Priority date: 2017-12-28
Filing date: 2017-12-28
Publication date: 2018-04-24

Abstract

The invention discloses a kind of NLP artificial intelligence approaches and interactive system based on financial vertical knowledge mapping, pass through the financial vertical collection of illustrative plates of foundation, it is associated with NLP, that is, natural language processing, so as to establish a set of perfect intelligent finance problem interactive system, with traditional search, question answering system is different by dynamic response, such a method can carry out active rhetorical question, recommend, excavate the problem of user's deep layer is wanted to ask, in flow, the problem of such a method receives user, first pass through NLP technologies, then semanteme is analyzed, retrieved by semanteme into financial vertical collection of illustrative plates, draw a circle to approve the scene domain of question answering, choose most suitable answer and be pushed to client, the problem of client can so being proposed, problem is better understood from reference to knowledge mapping, it can more accurately retrieve answer at the same time.

Description

NLP artificial intelligence method and interaction system based on financial vertical knowledge graph

Technical Field

The invention relates to the field of artificial intelligence dialogue systems and robot languages.

Background

At present, financial managers generally play a role of shopping guide in the purchasing process of financial products, and customers need continuous communication companions in the whole thinking decision process. Unlike traditional product sales, for financial products, purchasing is only a start and the subsequent process requires more communication to serve as a care companion. However, it is obviously impossible to provide communication in all weather by using manual service to ensure high-quality and efficient service. The rise of AI intelligence makes it possible to technically realize all-weather efficient intelligent service for customers.

In the prior art, the technologies relied on by the man-machine conversation system are mainly divided into the following two types:

the method is based on a question-answer library technology, namely, keywords in sentences are extracted to be matched with a large amount of sorted questions and answers. Many intelligent customer service adopt this kind of method to replace the cost of some manual customer service, and this kind of method disadvantage is that there is no correlation among the massive data, in addition if some sentence word orders are changed, the meaning is totally different, but the matched answer is probably unanimous;

the technology based on the search engine is to screen and answer the questions of the user by using the results returned by the search engine, and on one hand, the technology obtains low-quality responses, and on the other hand, the obtained answers are not controllable.

It can be seen that there are at least the following problems in the prior art:

1. the problems and corresponding answers of the traditional method are relatively independently stored, and all the problems are not related, so that the user problems cannot be effectively presumed through semantics;

2. the traditional method is more lateral to solving common user problems and problems before sale, and most personalized problems after sale cannot be answered accurately;

3. the traditional method is based on template matching, the problem richness is achieved by generalization of artificial mass problems, various questions need to be thought of in the same problem, and a large amount of time and energy are consumed.

Disclosure of Invention

Aiming at the problems in the prior art, the application provides a brand-new solution.

The application discloses an NLP (natural language processing) artificial intelligence method based on a financial vertical knowledge graph to realize intelligent financial problem interaction between a machine and a client, which specifically comprises the following steps:

s1, asking questions, and the client asks financial related questions;

s2, processing the natural language NLP, and processing the problems proposed by the customers through the NLP technology;

s3, semantic analysis and understanding, selecting proper phrases and keywords representing entities/relations according to the processed information for subsequent retrieval;

s4, information retrieval, namely retrieving information obtained through semantic analysis and understanding in the financial vertical map;

and S5, outputting the result, generating answer output according to the search result to answer the question of the client, and guiding and mining the question.

The financial vertical map is produced by combing and storing financial domain knowledge and the relationship between the financial domain knowledge through a database by utilizing a machine learning technology.

Further, the making of the financial vertical map comprises the following steps:

1) preparing related learning materials of the financial field;

2) learning the relevant learning data of the financial field in a semi-supervised mode by means of a machine learning technology;

3) combing the learned financial domain knowledge entities and the relations between the financial domain knowledge entities and storing the learned financial domain knowledge entities in a database to obtain a database;

4) semi-supervised maintenance is carried out on the relationship between the financial field knowledge entities, and synonyms representing the same relationship are aggregated in the entity range related to the relationship;

5) the picture description is added to common entities and used as a retrieval item for selection, extraction and use in the subsequent result output, so that the diversification of interactive answer forms can be improved.

When each entity concept is stored, other existing entities are associated, and finally, a financial vertical map with mutual relations among the entities is formed.

Further, the NLP natural language processing includes the following steps in chronological order:

s2.1, processing data, wherein the processing data comprises the processes of complex and simple conversion, Chinese word segmentation, part of speech tagging, data cleaning, syntax analysis, entity recognition and/or voice character conversion;

s2.2, classifying the data processed problems by the classification model from the three aspects of problem type, user behavior and emotion recognition;

s2.3, extracting information, wherein the information extraction is sequentially carried out according to the steps of extraction based on part of speech tagging, extraction based on semantic analysis and extraction based on speech piece analysis;

the extraction based on Part-of-Speech tagging is called PoS (Part-of-Speech) extraction for short, namely, the extraction is selected according to tagged nouns, verbs, adjectives or other parts-of-Speech;

the extraction based on semantic analysis selects a mode based on industry keywords or an entity list to extract and call information, and the calling method is a common statistical method known in the industry, such as chi-square, information gain, mutual information technology and the like.

And the sentence analysis is used for extracting and calling information in a dependency tree mode, wherein the dependency tree represents the dependency relationship among words in a sentence.

S2.4, completing information, detecting the deletion of sentence structures in the questions proposed by the customers, and completing the missing part;

s2.5 information queue: putting the extracted key information into a queue, removing the key information from the queue after more than 5 times of conversation, and simultaneously resolving and paraphrasing the indicative pronouns in the sentences.

In which conversation and queue specifically means that the key information of the latest conversation spoken by the user is put into the information queue, and the key information entered earlier (for example, 5 rounds ahead) is moved out of the queue. The queue information is used for controlling the semantic range of information retrieval and providing a basis for multi-round problem reply.

Further, the information retrieval comprises the following steps in sequence:

s4.1, classifying the questions, and defining the scene range of the answer of the questions according to the information obtained by semantic analysis and understanding, wherein the types of the questions comprise: special scene type questions, pre-or post-engagement questions, QA (question and answer type) questions.

And S4.2, searching the questions, searching the information obtained by semantic analysis and understanding in the financial vertical map within the range of the circled scene after the question types are determined to obtain required entities and/or relations, and generating answers according to word co-occurrence similarity, sentence pattern matching strategies and/or word sequence consideration modes for sending the answers to the clients.

Word co-occurrence similarity is a statistical-based model, and in a segment of language expression, several words often co-occur (i.e., co-occur) in the same sentence or the same paragraph, so that the words are considered to be related in meaning. The similarity calculation is carried out by adopting the method, and the entity searched by the semantics is calculated by adopting the method, so that the accuracy of the subsequent output result is improved.

The sentence pattern matching strategy has the effect that relevant entities required by the response are extracted through the information retrieval step, but the relevant entities are not a complete sentence. Through the sentence pattern matching strategy, the answer sentence pattern can be correspondingly selected through the known question sentence pattern, and the blank slot filling of the 'information completion' part is multiplexed to generate the answer.

The word order consideration is a supplement of the co-occurrence similarity, after the entities are selected, a group of similar entities is actually selected, the priority ranking of the group of entities is determined according to the matching degree of the searched group of entities and the questions, the higher priority is presented in the answers, and the lower priority is the alternative answers.

Further, the result output comprises outputting the generated answer, the guided and excavated question to the customer, wherein the guided and excavated question is further guided and excavated according to the entity associated with the answer in the financial vertical map, and forms a new question to be output to the customer for the customer to answer or confirm, so that a more real and effective dialogue interaction state can be realized.

In addition, the application also discloses a system for carrying out intelligent financial problem interaction by using the NLP artificial intelligence method based on the financial vertical knowledge graph, which specifically comprises an input module, an NLP natural language processing module, a semantic analysis and understanding module, an information retrieval module and a result output module;

wherein,

the input module is used for receiving financial related questions put forward by a customer and sending the questions to the NLP natural language processing module for processing;

the NLP natural language processing module processes the problems proposed by the customers through an NLP technology;

the semantic analysis and understanding module selects proper phrases and keywords representing entities/relations according to the processed information for subsequent retrieval;

the information retrieval module retrieves the information obtained by semantic analysis and understanding in the financial vertical map;

and the result output module generates answer output according to the retrieval result so as to answer the question of the client, and guides and mines the question.

And the established financial vertical map is associated with NLP (natural language processing), so that a set of complete intelligent financial problem interaction system is established. Different from the passive response of the traditional searching and question-answering system, the method can actively ask a question, recommend and mine questions deeply asked by the user. In the process, the method receives the questions of the user, firstly, the NLP technology is used, then, the semantics are analyzed, the financial vertical map is searched through the semantics, the scene range of question answers is defined, and the most suitable answers are selected and pushed to the client.

By doing so, on one hand, the questions proposed by the customers can be better understood by combining the knowledge graph, and on the other hand, the answers can be more accurately retrieved. A plurality of knowledge points exist in the whole financial activity, the relation of the related knowledge points is combed and appropriately stored by a graph database for subsequent reading and retrieval of an NLP system.

Compared with the prior art, the application has the advantages that:

1. the established financial knowledge map can continuously perfect and enrich entity/relation data in the database, when each entity concept is stored, other existing entities can be intelligently associated, and a financial vertical map with complex relation is finally formed, so that complex personalized problems of customers can be solved, and meanwhile, the diversification of interactive answer forms can be improved by storing different types of data (such as picture information) and a special information retrieval and answer generation output mechanism.

2. Active question asking can be performed, questions deeply asked by the user can be recommended and mined, and the more real and effective conversation interaction state is realized.

3. The client problems can be better and more efficiently understood, the most appropriate answers are accurately retrieved, and the conversation communication efficiency is greatly improved.

Drawings

FIG. 1 is an illustration of an example of a converged knowledge entity in a financial vertical graph stored in a graph database;

FIG. 2 is a general flow diagram of an intelligent dialog;

FIG. 3 is a detailed flow diagram of a financial vertical knowledge-graph based NLP artificial intelligence method to achieve intelligent financial problem interaction between machines and customers.

Detailed Description

For the purpose of full disclosure, the present invention will be described in further detail with reference to the following examples. It should be understood that the specific examples described below are only for illustrating the present invention and are not intended to limit the scope of the present invention.

The application specifically discloses an NLP artificial intelligence method based on a financial vertical knowledge graph to realize intelligent financial problem interaction between a machine and a client, and the method specifically comprises the following steps:

s1, asking questions, and the client asks financial related questions;

An example of a financial knowledge entity in a financial vertical as shown in FIG. 1, the production of the financial vertical comprising:

1) preparing financial field-related learning materials, which are materials of bank-related knowledge in the example shown in fig. 1;

2) learning the relevant learning materials of the financial field by using a semi-supervised mode by means of a machine learning technology, wherein in the example, the learning materials contain information of four major lines of China and relevant relations thereof;

3) the learned financial field knowledge entities and the relations between the financial field knowledge entities are combed and stored in a database to obtain a graph database, in the example, learning that 'four major banks in China have China Bank, China Industrial Bank, China agricultural Bank and China construction Bank', 5 entities and 4 groups of relations are learned through the sentence, namely four major banks in China-having-China Bank, China Industrial Bank in China-having-China, four major banks in China-having-China agricultural Bank in China-having-China Bank in China and four major banks in China-having-China construction Bank in China;

4) and performing semi-supervised maintenance on the relationship between the knowledge entities in the financial field, and aggregating synonyms representing the same relationship in the entity range related to the relationship, for example, receiving the financial knowledge again, wherein the four major banks in China comprise a Chinese bank, a Chinese industrial and commercial bank, a Chinese agricultural bank and a Chinese construction bank. New relations including appear among 5 existing entities again, and two semantically existing relations including and including are consistent, can be combined into a relation, and records close relations to the entities, so that the accuracy of machine learning is improved;

5) adding picture description to common entities, using the picture description as a retrieval item for subsequent selection and extraction, and importing related pictures to the common entities for selection and extraction during reply in order to improve the diversity of interactive answer forms by storing the pictures.

As shown in fig. 3, fig. 3 shows a detailed flow of NLP artificial intelligence method based on financial vertical knowledge graph to realize intelligent financial problem interaction between machine and customer,

the NLP natural language processing in step S2 includes the following steps in order:

the mathematical expression of the method adopted by Chinese word segmentation is as follows:

(T)＝(W₁)P(W₂|W₁)P(W₃|W₂)…P(W_n|W_n-1)

p represents the probability of occurrence, where W₁Represents the result of the first word region recognition, so P (W)₁) Representing the probability of the first word occurring. P (W)₂|W₁) Representing the probability of a second word appearing if the first word appears, W can be seen₁W₂Whether it is a common word. Thus, a sentence with divided words is generated until the end of the sentence. Such as: china | four great | banks | which |?

the question type is mainly the question to distinguish the user is the 5w type question (where, what, while, why, how), the regular question or the non-question type communication.

User behavior refers to the question that the user makes on which page, under what circumstances.

And the emotion recognition can extract the related emotion words according to the conversation process of the user.

More specifically, the mathematical expression in which information extraction is performed by the chi-square:

the feature extraction algorithm is divided into two categories of feature selection and feature extraction. The chi-square test belongs to the better algorithm in the feature selection algorithm.

t and c are two random variables, chi 2 represents chi-square value for checking the correlation of data, t represents a word, and c represents a category. For example, t may represent a flower and c may represent a plant.

When we find N articles, t is classified as containing and not containing, and c is classified as belonging and not belonging. So that the following tables can be generated, corresponding to the letters in the above formulas, respectively.

Feature selection	c1. Belonging to the genus of "plants"	c2. Not belonging to "plants"	Total of
				t1. comprising "flowers"	A	B	A+B
t2. not containing "flowers"	C	D	C+D
				Total number of	A+C	B+D	N

The mathematical expression in which information extraction is performed by information gain:

in the case where the feature is that Y is fixed, the conditional entropy of X is H (X | Y), and P (X | Y) is the probability of occurrence. The colloquial explanation is how much the uncertainty of X is reduced after knowing Y, compared to when it is not. The two variables X, Y of the information gain differ in position, which is a means to consider Y as reducing X uncertainty. While the two variables of mutual information are in the same place.

The mathematical expression for extracting information through mutual information is as follows:

g(D，A)＝H(D)-H(D|A)

h (D) represents the self information quantity, i.e. the sent information, H (D | A) represents the condition information quantity, and g (D, A) represents the mutual information quantity. The mutual information amount is self information amount-condition information amount. Information extraction is carried out through mutual information, and the internal bonding strength of the character strings is calculated by utilizing the mutual information, so that the extracted information is more complete and is not the state of the keyword. Such as natural language processing, which finds these words as a fixed academic vocabulary through mutual information, and therefore does not break down into three words for understanding.

S2.4, completing information, namely detecting the deletion of a sentence structure in a question proposed by a client (also called slot blank detection), and completing the missing part (also called blank slot filling);

Wherein the information retrieval in step S4 includes the following steps in order of sequence:

Co-occurrence similarity calculation mathematical expression:

Similarity(s1,s2)＝αSDMG(s1,s2)+βSDMG(s1,s2)+γSDMG(s1,s2)

s1, s2 are two words, α, β, γ are co-occurrence correlations in three respective spatial vectors.

Wherein the outputting of the result of step S4 includes outputting the generated answer, guided and mined question to the customer, wherein the guided and mined question is further guided and mined according to the entity associated with the answer in the financial vertical map to form a new question to be output to the customer for the customer to answer or confirm, so that a more real and effective dialogue interaction state can be realized. For example:

the end user asks: which are the four banks in china?

And (3) dialog answers: the four banks in China all have China Bank, China Industrial and commercial Bank, China agricultural Bank, China construction Bank, do you know? Wherein the bank of the chinese industry and commerce was established in 1984.

The embodiment can be specifically realized by a system for intelligent financial problem interaction by using an NLP artificial intelligence method based on a financial vertical knowledge graph, wherein the system comprises an input module, an NLP natural language processing module, a semantic analysis and understanding module, an information retrieval module and a result output module;

wherein,

The above-described example embodiments merely represent embodiments of the present patent, and the description thereof should not be construed as limiting the scope of the patent. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the patent idea, which falls within the protection scope of the present invention. Therefore, the protection scope of this patent shall be subject to the appended claims. The above-mentioned embodiments only express the embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. An NLP artificial intelligence method based on a financial vertical knowledge graph is characterized by comprising the following steps:

s1, asking questions, and the client asks financial related questions;

s3, semantic analysis and understanding, wherein phrases and keywords representing entities/relations are selected according to the processed information for subsequent retrieval;

s5, outputting results, generating answers according to the search results to output answers to the questions of the clients, and guiding and mining the questions;

2. The NLP artificial intelligence method based on financial vertical knowledge-graph as claimed in claim 1, wherein the making of the financial vertical knowledge-graph comprises:

1) preparing related learning materials of the financial field;

5) adding picture description to common entities, and using the picture description as a retrieval item for selection, extraction and use when outputting subsequent results, thereby improving the diversification of interactive answer forms;

3. The NLP artificial intelligence method based on financial vertical knowledge-graph as claimed in claim 1, wherein the NLP natural language processing comprises the following steps in order of precedence:

4. The NLP artificial intelligence method based on financial vertical knowledge-graph as claimed in claim 1, wherein said information retrieval comprises the following steps in the order of precedence:

s4.1, classifying the questions, and defining the scene range of the answer of the questions according to the information obtained by semantic analysis and understanding, wherein the types of the questions comprise: special scene questions, pre-or post-engagement questions, QA (question and answer) questions;

5. The NLP artificial intelligence method based on financial vertical knowledge graph as claimed in claim 1, wherein the result output includes outputting the generated answers, guided and mined questions to the customer, wherein the guided and mined questions are further guided and mined according to the entity associated with the answers in the financial vertical knowledge graph to form new questions to be output to the customer for the customer to answer or confirm.

6. A system for intelligent financial problem interaction by using the NLP artificial intelligence method based on financial vertical knowledge graph according to any one of claims 1-5, wherein the system comprises an input module, an NLP natural language processing module, a semantic analysis and understanding module, an information retrieval module and a result output module;

wherein,