CN118132719A - Intelligent dialogue method and system based on natural language processing - Google Patents

Abstract

The invention discloses an intelligent dialogue method based on natural language processing, which comprises the following steps: semantic parsing and intent recognition; intent matching; emotion analysis; searching a knowledge base; a multi-modal answer; modeling a customer portrait; machine learning. The invention is based on natural language understanding technology construction, comprehensively applies a plurality of artificial intelligence technologies such as machine learning, knowledge graph and the like, and has the functions of chat consultation, semantic association, intention recommendation, customer portraits and the like. The invention realizes the aggregation of financial service capability, meets the data use requirement of one-stop and cross-platform, and can promote the usability of the system and the perception degree of the financial intelligent application of the user; the system can simply, quickly and effectively serve personnel at each level, realize the accurate query and intelligent management of knowledge such as business processes, system terms, system operation and the like, and promote business consultation to be converted into on-line, unmanned and intelligent; meanwhile, the investment of the operation and maintenance cost of enterprises is reduced.

Description

Intelligent dialogue method and system based on natural language processing

Technical Field

The invention relates to the technical field of artificial intelligence natural language processing, in particular to an intelligent dialogue method and system based on natural language processing.

Background

The financial management is provided with a plurality of sets of financial information systems such as ERP, financial management and control and the like, when financial staff applies various system tools in daily work, the problems of business process class, system operation class and function error reporting class are solved, the operation staff is required to be searched for processing, or the work order submitting mode is required to wait for the operation staff to process, and the time consumption and the efficiency are low; when related system bases are needed to be searched when problems are encountered in business processing, key points are needed to be manually read from a large number of system files, and convenience is low; when data is required to be queried so as to be convenient for analysis or report preparation, different types of data need to be queried across a plurality of systems, and man-machine interaction is inconvenient; the problems of inconvenient system use, untimely response and the like exist for financial staff; a number of problems for customer service are simply repetitive problems and do not form an effective knowledge accumulation. Along with the rapid development of various artificial intelligence technologies such as natural language processing, machine learning, knowledge graph and the like, the construction of a diversified and more intelligent man-machine interaction mode becomes easier to realize.

Natural language processing (Natural Language Processing) is a discipline that studies interactions between human language and computers in order to enable computers to understand, process and generate natural language. NLP technology relates to a number of fields including language understanding, language generation, machine translation, information retrieval, text classification, emotion analysis, etc. (1) In terms of language understanding, NLP techniques can help computing mechanisms to solve the meaning and grammatical structure of human language. This includes tasks such as lexical analysis (breaking up sentences into words), syntactic analysis (analyzing the syntactic structure of sentences), semantic analysis (understanding the meaning of sentences), and phonetic analysis (understanding the meaning of sentences in context); (2) In terms of language generation, NLP techniques can help computers generate natural language text that conforms to grammatical and semantic rules. This includes text generation, machine translation, automatic summarization, etc.; (3) In terms of machine translation, NLP technology can automatically translate text in one language into another. This involves word sense alignment, syntactic structure conversion, semantic conversion, etc.; (4) In terms of information retrieval, NLP technology can help computers retrieve information relevant to a user query from a large volume of text. The method comprises the technologies of keyword matching, text similarity calculation, text classification and the like; (5) In emotion analysis, NLP techniques can help computers understand and analyze emotion tendencies in text. This may be used for emotion analysis, emotion recommendation, etc. applications. The core of NLP technology is to build models and algorithms to process natural language. Common techniques include statistical models, machine learning, deep learning, and the like. Statistical models learn the rules and patterns of language by analyzing a large corpus. Machine learning techniques automatically recognize and classify text by training models. Deep learning techniques simulate the human language processing process by constructing deep neural networks.

Natural language understanding (Natural Language Understanding) is an important research direction in the field of artificial intelligence, aimed at enabling computers to understand and process the meaning and semantics of human natural language. The method mainly comprises five aspects of grammar analysis, semantic analysis, intention recognition, context understanding and emotion analysis: (1) syntax analysis: is a process of decomposing a natural language sentence into grammatical structures. It can identify individual components of sentences (e.g., subject, predicate, object, etc.) and determine relationships between them, common parsing methods include rule-based methods and statistical-based methods; (2) semantic analysis: is the process of understanding the meaning and semantics of sentences. It can identify entities, relationships, and events in sentences and convert them into a form understandable by a computer. Common semantic analysis methods include word sense disambiguation, entity identification, relationship extraction, etc.; (3) intention recognition: is a process of identifying the user's intent and needs in a conversation. It can determine what the user wants to do based on the user's input and translate it into a corresponding action or answer. Common intent recognition methods include rule-based methods and machine learning-based methods; (4) contextual understanding: is a process of understanding context information in a conversation. It can understand the current dialog from the previous dialog content and respond or operate accordingly based on the context information. Common context-aware methods include rule-based methods and memory-based methods; (5) emotion analysis: is a process of analyzing emotional tendency and emotional state in a text. It can recognize emotion words in text and judge emotion polarity (such as positive, negative, neutral, etc.). Common emotion analysis methods include rule-based methods and machine learning-based methods.

Machine learning (MACHINE LEARNING) is a branch of artificial intelligence that aims to enable computers to automatically learn and improve performance through data and experience without explicit programming. It builds models using statistical and optimization algorithms, so that the computer can learn from the data and make predictions or decisions based on the learned knowledge. Machine learning techniques can be divided into three main types, supervised learning, unsupervised learning, and reinforcement learning. (1) supervised learning: in supervised learning, the computer learns from labeled training data. These tags provide correspondence between the input data and the desired output. By learning these correspondences, the computer can make predictions from the new input data. Common supervised learning algorithms include linear regression, logistic regression, decision trees, support vector machines, and neural networks; (2) unsupervised learning: in unsupervised learning, the computer learns from training data without tags. It attempts to find patterns and structures in the data to cluster, dimension down, or anomaly detect the data. Common unsupervised learning algorithms include clustering algorithms (e.g., K-means clustering and hierarchical clustering), association rule mining, and principal component analysis; (3) reinforcement learning: in reinforcement learning, a computer learns by interacting with the environment. It learns how to take action by way of trial and error to maximize the expected rewards. Reinforcement learning is commonly used in the fields of games, robot control, and autopilot.

Disclosure of Invention

The invention aims to provide an intelligent dialogue method and system based on natural language processing, which realize five service capacities of financial information inquiry, financial system search, system operation inquiry, professional dictionary interpretation, boring and the like by constructing a professional knowledge base of power grid business; through semantic understanding and intention recognition, accurate positioning user demands, all-weather, delay-free and personalized financial professional consultation services are provided for company staff, and the pressure of financial operation and maintenance staff is reduced to a certain extent, so that the enterprise operation and maintenance cost investment is reduced.

In order to achieve the above purpose, the technical scheme of the invention is as follows:

an intelligent dialogue method based on natural language processing comprises the following steps:

s1: semantic parsing and intent recognition;

Judging the intention of a user according to the text input by the user by adopting a mode of combining a text retrieval mode, a natural sentence pattern matching mode, a rule sentence pattern matching mode and a meaning type matching shallow semantic understanding mode with deep learning mode, and understanding the requirement of the user; the system has a semantic association function, a user can input fuzzy semantics, and the system automatically brings up maintained expansion questions to guide the user to accurately express intention;

S2: intent matching.

According to the S1 intention recognition result, similarity matching is carried out between the user problem and the similarity condition of the knowledge points in the system knowledge base in the high-dimensional semantic space by using sentence vectors;

s3: and (5) emotion analysis.

According to S1 semantic analysis results, through 30 emotion models such as happiness, excitement, gas generation and the like in the system, the emotion of the user is perceived in real time, and corresponding answers are made according to the emotion states of the user;

S4: and searching a knowledge base.

According to the results of S2 and S3, firstly extracting keywords, then searching a knowledge base by adopting a word vector model and a semantic similarity calculation method, outputting a text with highest similarity, and simultaneously, moisturizing the text by utilizing an emotion model;

s5: multiple forms of answers.

Generating an answer to the S4 matched information according to predefined rules and templates, wherein the content forms comprise but are not limited to characters, documents, tables, maps, links, videos and the like; for example: for the data query class, the data query class can be replied in the form of text and documents; aiming at financial system and system operation consultation class, the system can reply in the form of text, picture and file. Meanwhile, the system has a function of 'related questions', and the user can independently click on the 'related questions' according to the analysis of the user questions while outputting the answers, so that the user can be helped to more comprehensively and deeply learn about the consultation questions.

S6: customer representation modeling.

The system has the capability of counting, analyzing and classifying the problems, automatically generates customer portraits according to the consultation records of the users, and displays 'guessing you want to ask', 'hot spot problems', 'my portraits' on a dialogue interface, so that the users can more conveniently and quickly know the business services related to the customer portraits.

S7: machine learning.

The system provides channels for demand collection and user feedback in a plurality of functional modules such as interactive windows, conversation processes, system settings and the like. The system is internally provided with a machine learning platform, so that algorithm and model level tuning can be realized, and the answer and response capability of the system can be continuously optimized according to the machine learning algorithm.

In the invention S1, the system built-in NLP engine is used for configuring parameters such as threshold, weight, inheritance meaning class and the like, so that the accuracy degree of identifying the user intention by the system is improved. The system adopts a mode of combining a shallow semantic understanding mode such as text retrieval, natural sentence pattern matching, regular sentence pattern matching and the like with deep learning, and calculates the similarity condition of a user problem and a knowledge point in a knowledge base in a high-dimensional semantic space by using sentence vectors, so as to further judge the intention of the user. Through adding the semantic association menu function on the dialogue interface, the user can input fuzzy semantics, the system automatically brings out maintained expansion questions, guides the user to accurately express intention, and can accurately position the problems and provide correct answers, thereby improving the practicability of the system.

In S1, the system is displayed on the desktop of the user by three-dimensional robot images (including but not limited to an out-of-field image, a standing and standby image, a thank you image, a sleeping image, a jumping image and an off-field image), the user can enter the consultation dialogue interface by double-clicking the robot images, a convenient inquiry mode is provided for the user, and the traditional help-transfer belt mode of 'mouth-mouth phase transfer' is changed.

In S2, the system has a menu authority control function, and only the user with the menu authority can perform corresponding operations (including but not limited to viewing and downloading). After S1 intention recognition, the system checks whether the user has the corresponding use authority of the consultation report, and if the authority check is passed, the system replies correspondingly; if the permission verification is not passed, the system replies: to the best, you do not have the related XX use rights, please contact the XX system personnel to apply.

In S3, 30+ emotion models such as happiness, excitement, gas generation and the like are built in the system, and a user-defined emotion model is supported, so that the emotion of a customer can be perceived in real time, and emotion soothing is actively carried out in a dialogue, so that the customer experience is greatly improved.

And S4, extracting keywords according to the results of the S2 and the S3, searching a knowledge base by adopting a word vector model and a semantic similarity calculation method, outputting a text with highest similarity, and coloring the text by utilizing an emotion model.

S4, the system constructs a huge knowledge base through a knowledge graph technology, wherein the knowledge base comprises knowledge and information in the financial field and the office field. The knowledge base provides a plurality of modes such as uploading and the like for updating the domain resources, and has an intention checking function, if the following conditions exist, the core words are not imported. Including but not limited to (1) intended name repeat or excess length; (2) sentence pattern disagreement and rule or sentence pattern repetition; (3) Only importing corresponding intentions and expansion questions when the intended word slots do not exist; (4) The action keywords and the business keywords are not intended to be imported when the action keywords and the business keywords are not present; (5) The intent scene name already exists and the intent is automatically merged with the existing scene.

In S5, the system supports collection of a plurality of slot values, inheritance slot values and verification slot values in a plurality of rounds of conversations, and has the capability of extracting and inheriting the multi-slot values in the same conversation process, so that the conversations are more intelligent.

In S6, the system automatically sorts the hot spot problems according to the analysis result of the consultation frequency, and the user can realize quick consultation through the hot spot problem list.

The working principle of the invention is as follows: the system adopts a mode of combining shallow semantic understanding and deep learning to carry out semantic understanding and intention recognition on user input information, and then combines an analysis result of an emotion model to carry out similarity matching on a knowledge base so as to provide a most relevant answer for a user; the system has the capability of counting, analyzing and classifying the problems, and automatically generates customer portraits according to the consultation records of the users; the system can realize algorithm and model level tuning through a built-in machine learning platform.

The method adopts a mode of combining a shallow semantic understanding mode such as text retrieval, natural sentence pattern matching, regular sentence pattern matching and the like with deep learning, calculates the similarity condition of a user problem and a knowledge point in a knowledge base in a high-dimensional semantic space by using sentence vectors, and further judges the intention of the user. The accurate answer is given to the user through the multi-round dialogue engine and the emotion model which are arranged in the dialogue system, the business consultation requirement of the user in one-stop and cross-platform mode is met, and the innovation and the value creation of the financial management are assisted.

The beneficial effects of the invention are as follows:

1. Provide convenient financial man-machine interaction: the system provides services such as financial consultation, lease contract information extraction and creation, 7 x 24 data inquiry, enterprise system inquiry, system operation consultation, financial dictionary solution and the like based on natural language by establishing a financial knowledge base, integrating data center, integrating various business system services and the like in the form of PC client application. Financial staff can input characters on line by self or conduct intelligent consultation in the form of language dialogue without seeking help of professional operation staff. Aiming at various problems, intelligent experts search for similar conditions of semantic environments by adopting text retrieval, natural sentence patterns, regular sentence pattern matching and other modes, further judge the intention of users, perform quick and accurate response on personalized inquiry, realize automatic analysis and automatic input of various knowledge information by adopting an intelligent knowledge analysis and reading and learning mode, and reduce the manual workload by 70%.

2. Providing various types of financial accounting services: and aiming at consultation of the system specification type, providing quick query of the system requirement. Multiple rounds of interaction are supported, context meaning is understood, system document downloading is supported, and related problems are automatically recommended according to user problems. Machine learning is performed to address the need for consultation in different semantic contexts, providing accurate responses, rather than just providing fuzzy queries as in search tools. Aiming at the consultation of the system application, the method provides accurate positioning and quick inquiry of the application problems, supports the transverse browsing of the problems of the process operation, supports the one-key call of related services in the middle station of the company, and rapidly responds to the problems in the aspects of flow, operation and the like in the daily transaction processing and the like of financial staff. Aiming at core index type consultation, single consultation and multiple rounds of interaction are supported, diversified consultation sentence patterns and content guiding consultation are supported, common indexes are automatically pushed, and different query contents are displayed and guided based on authority management of roles. Providing dynamic information prediction support, providing supplemental presentation of information that may be relevant to the problem being consulted, such as: for the tax rate query problem, on the basis of providing tax rate data, policy information relevant to tax processing and the like are additionally provided.

3. Remodelling the system operation and maintenance service system: the system is responsible for solving the high-frequency conventional problems of staff in all weather, and in the service providing process, the self function of machine learning continuous iteration is utilized to simulate human learning behaviors. Along with the continuous accumulation of chat records of users, the system automatically analyzes and generalizes the problems to form new financial knowledge points, continuously expands a financial knowledge base, continuously accumulates new knowledge and new skills without boundaries in the service process, and realizes unconscious knowledge accumulation and conscious knowledge inheritance while continuously expanding the radius of consultation service. Traditional operation and maintenance personnel focus on answering the unresolved questions of the machine and continuously help the machine to train, learn and expand the knowledge base. The man-machine cooperation mode can greatly improve operation and maintenance service efficiency and response speed, improve the learning times of financial staff on financial tax system, and continuously reduce financial information operation and maintenance management cost.

Drawings

Fig. 1 is a schematic diagram of the present invention.

Fig. 2 is a deep learning model diagram in the present invention.

Fig. 3 is a diagram of a machine learning model in the present invention.

Detailed Description

The following describes the embodiments of the present invention further with reference to the drawings and examples. The following examples are only for more clearly illustrating the technical aspects of the present invention, and are not intended to limit the scope of the present invention.

Fig. 1 is a schematic diagram of the present invention, which includes the following steps:

S1, semantic parsing and intention recognition: the system judges the intention of the user and understands the requirement of the user according to the text input by the user by adopting a mode of combining a shallow semantic understanding mode such as text retrieval, natural sentence pattern matching, regular sentence pattern matching, meaning class matching and the like with deep learning;

shallow semantic understanding: the method comprises text retrieval, natural sentence pattern matching, rule sentence pattern matching and meaning class matching.

1. Text retrieval: the Lucene algorithm is selected and used as an open-source full-text search engine, and the score of an input text and a text library can be calculated according to similar words;

2. natural sentence pattern matching: namely full text matching;

3. rule sentence pattern matching: searching whether a sentence accords with a predesigned rule sentence pattern or not through fuzzy matching of keywords or keyword groups;

4. And (5) meaning class matching: the user classifies the words and gives weights, and the similarity of the two sentences is calculated according to the meaning attribute of the words designed by the user, so as to give a score.

Deep semantic understanding: and calculating the similarity of the meaning expressions of the two sentences according to the deep learning model, providing a semantic association function for the scoring system, inputting fuzzy semantics by a user, automatically taking out maintained expansion questions by the system, and guiding the user to accurately express the intention. Fig. 2 is a deep learning model diagram in the present invention.

S2, intention matching: according to the S1 intention recognition result, the user problem and the knowledge points in the knowledge base are expressed as high-dimensional semantic vectors, cosine similarity matching calculation is carried out by using a sentence vector model, the similarity condition of the user problem and the knowledge points in the system knowledge base in the high-dimensional semantic space is known, and the knowledge points most relevant to the user problem are found, so that more accurate answers or suggestions are provided.

The steps for calculating the similarity of two sentences are as follows:

1. through Chinese word segmentation, the complete sentence is divided into independent word sets according to word segmentation algorithm

2. Find the union of two word sets (word bag)

3. Calculating word frequencies of respective word sets and vectorizing the word frequencies

4. The text similarity can be obtained by carrying the vector calculation model

The cosine similarity value can be clearly calculated through a calculation model formula:

Vector a (x 1, x2, x3, x4, x 5), vector b (y 1, y2, y3, y4, y 5);

The molecules are (x 1y 1) + (x 2y 2) + (x 3y 3) + (x 4y 4) + (x 5y 5)

The denominator is sqrt (x1x1+x2x2+x3x3+x4x4+x5x5)

S3, emotion analysis: according to the S1 semantic analysis result, emotion analysis is carried out through 30 emotion analysis models such as happiness, excitement, gas generation and the like in the system, the emotion of the user is perceived in real time, and corresponding answers are made according to the emotion state of the user. The emotion analysis model employed by the system is a machine learning model that recognizes and classifies emotion tendencies, such as positive, negative or neutral, in text by learning features and patterns in the text and provides an emotion score for each text. The system trains the emotion analysis model using a large amount of marking data, where each text is marked as positive, negative or neutral. The model predicts the emotion of the new text by learning patterns and features in these markup data. FIG. 3 is a diagram of a machine learning model in the present invention, the machine learning model is implemented as follows:

in training process (a), model learning associates specific inputs (i.e., text) with corresponding outputs (tags) based on training samples. The feature extractor transmits the text input into a feature vector. Pairs of feature vectors and labels (e.g., positive, negative, or neutral) are fed into a machine learning algorithm to generate a model.

In the prediction process (b), a feature extractor is used to transform the unseen text input into feature vectors. These feature vectors are then fed to a model that generates predictive labels (positive, negative or neutral).

S4, searching a knowledge base: according to the results of S2 and S3, firstly extracting keywords, then searching a knowledge base by adopting a word vector model and a semantic similarity calculation method, outputting a text with highest similarity, and simultaneously, moisturizing the text by utilizing an emotion model;

The system constructs a huge knowledge base containing knowledge and information in the financial field and the office field through the knowledge graph, so that the knowledge base can be queried to answer the problems of the user, and accurate and detailed information is provided.

1. Knowledge graph construction: including data acquisition and data cleaning and processing. Data collection is the collection of structured and semi-structured data, including entity, relationship, and attribute information, from a variety of sources (e.g., the Internet, PDF documents, word documents, text documents, databases); data cleaning and processing is to clean, deduplicate and format the collected data to ensure consistency and accuracy of the data.

2. Knowledge representation: the information in the knowledge-graph is represented as a combination of entities, relationships and attributes using an entity-relationship-attribute model, while the knowledge-graph data is efficiently stored and queried using a graph database and a triplet storage system.

3. Entity identification and linking: natural language processing techniques are used to identify entities in text, such as names of people, places, organizations, etc., and link the entities in text with entities in knowledge-graph to establish semantic associations.

4. And (3) relation extraction: NLP techniques are used to extract relationships between entities from text, such as "parent-child relationships" or "work on relationships," while modeling the extracted relationships as edges in a graph.

5. Updating and maintaining a knowledge graph: the knowledge graph is updated periodically, and errors and inconsistencies in the knowledge graph are monitored and corrected to reflect changes in the real world and growth of new knowledge.

6. Inquiring the knowledge graph: information in the knowledge graph is retrieved using a query language such as SPARQL, and complex graph query operations are performed using a specialized graph database query language.

7. Knowledge graph reasoning: semantic reasoning is performed based on the relationships and attributes in the knowledge graph, and a rule logic reasoning engine is used for performing reasoning operation, so that knowledge reasoning is performed.

8. Security and privacy: protecting the data in the knowledge graph, and ensuring that the data is not accessed and abused by unauthorized; privacy protection measures such as data desensitization and authority control are taken when sensitive information in the knowledge graph is processed.

S5 multimodal answer: generating an answer to the S4 matched information according to predefined rules and templates, wherein the content forms comprise but are not limited to characters, documents, tables, maps, links, videos and the like; for example: for the data query class, the data query class can be replied in the form of text and documents; aiming at financial system and system operation consultation class, the system can reply in the form of text, picture and file. Meanwhile, the system has a function of 'related questions', and the user can independently click on the 'related questions' according to the analysis of the user questions while outputting the answers, so that the user can be helped to more comprehensively and deeply learn about the consultation questions.

In addition, the system can carry out multiple rounds of conversations with the user, can understand the language of the person like a person, can carry out multiple rounds of questions and answers, and grab valuable available word slots in the questions and answers process to carry out next round of questions and answers or carry out intelligent topic transfer.

1. Dialog state tracking: the dialog states in the dialog process, including the user's current intent, contextual information, and the last reply of the robot, are tracked, and updated and maintained using a state transition model and rule-based methods.

2. And (5) intention recognition: the user's intent, i.e., the back purpose and demand of the user input, is identified. Intent classification may be performed using techniques such as natural language processing and machine learning to determine the behavior desired by the user.

3. Reply generation: and generating a reply of the robot according to the dialogue state and the intention of the user. Appropriate replies may be generated from the dialog history and context using template matching, natural language generation models, or rule-based methods.

4. Dialogue decision: based on the dialog state and the goals of the robot, a decision is made as to how to handle the user's input. Using reinforcement learning methods, dialogue questions are modeled as a markov decision process (Markov Decision Process, MDP) and replies are selected by learning the best strategy.

5. Context management: in a multi-round conversation, the contextual information of the conversation is managed, ensuring that the system is able to understand and use the previous conversation history. A memory network may be used to store and retrieve dialog history, as well as to correlate current dialog and context.

6. Error handling: errors that may occur in the dialog are handled, such as ambiguities or incompleteness of the user input, or errors in the understanding of the robot. This may correct errors and provide a more accurate reply by using confidence metrics of the model or adding feedback mechanisms.

7. Meaning understanding and supplementary questioning: in a dialogue, when the robot cannot understand the user's input or needs more information, the system will make meaning understanding and supplementary questions, identify ambiguous or missing information of the user using rule-based or machine-learning-based methods, and generate appropriate queries for clarification.

8. Intelligent topic transfer: in the dialogue, topic transition and transition are performed according to user input. Techniques may be used to detect changes in user intent and to shift topics according to established dialog strategies. These technical details are some basic elements of constructing dialog management in an intelligent artificial chat robot, and specific implementation and algorithm selection may vary depending on application scenarios and requirements.

S6, modeling a customer portrait: the system has the capability of counting, analyzing and classifying the problems, automatically generates customer portraits according to the consultation records of the users, and displays 'guessing you want to ask', 'hot spot problems', 'my portraits' on a dialogue interface, so that the users can more conveniently and quickly know the business services related to the customer portraits. Customer representation modeling techniques are methods for creating customer representations by collecting, sorting, and analyzing customer data. A customer representation is a comprehensive description of a customer, including the customer's basic information, behavioral characteristics, preferences, and requirements, among others. Through customer portrait, the system can better know the customers, provide personalized information and services, and improve the satisfaction degree and loyalty of the customers to the system. The process of customer representation modeling is as follows:

1. Data collection and arrangement: customer operation logs and other various information are collected through system functions, including basic information, behavior data, feedback data and the like. Then cleaning, sorting and classifying the data for subsequent analysis and modeling;

2. Data analysis and mining: the data analysis and mining technology is utilized to count, analyze and mine the customer data, and discover the behavior patterns, preferences, demands and the like of the customers. Common techniques include cluster analysis, association rule mining, predictive modeling, etc.;

3. machine learning and artificial intelligence: model training and predicting the client data by utilizing machine learning and artificial intelligence technology;

4. visualization and presentation: the customer portrait is displayed in the form of window text and automatically adjusted according to the real-time data of the user.

S7, machine learning: the system provides channels for demand collection and user feedback in a plurality of functional modules such as interactive windows, conversation processes, system settings and the like. The system is internally provided with a machine learning platform, can realize algorithm and model level tuning, continuously optimizes own answer and response capacity according to the machine learning algorithm, and improves own expression capacity through a large amount of dialogue data, so that the answer is more ready and smooth.

1. Question-answering model: the model uses techniques similar to the attention mechanism to extract information across text paragraphs, enabling answers to questions posed by a user from a given text paragraph.

2. Intent recognition and semantic understanding: the text of the user is classified to determine the user's intent and key information, such as date, place, name, etc., is extracted from the text input by the user, and semantic roles of the individual components in the sentence, such as subject, predicate, object, etc., are analyzed.

3. Dialog generation and dialog management: using the sequence-to-sequence model for generating a response to the chat, including a recurrent neural network, a transducer model, and an attention mechanism; using dialog state tracking to track the state of dialog to determine an appropriate answer; a dialogue strategy is used to determine how and when to answer a user's question to provide a meaningful reply.

4. Model training and optimization: the model is trained using a large amount of annotation data, including dialogue records collected from the internet, and fine-tuned by migration learning to adapt to the tasks and fields of the system, and finally, evaluation and tuning of the model performance is performed using evaluation metrics and techniques.

5. Real-time and performance optimization: in order to provide fast response in real-time conversations, models are optimized, including model compression, acceleration, and deployment on high performance hardware.

6. Continuous improvement and user feedback: user feedback and coverage are collected to improve the performance of the chat robot. Reinforcement learning techniques are used, such as reinforcement dialog strategies, model updates using user feedback, and the like.

According to the invention, five service capacities of financial information inquiry, financial system search, system operation inquiry, professional dictionary interpretation, boring and the like are realized by constructing the professional knowledge base of the power grid business; through semantic understanding and intention recognition, accurate positioning user demands, all-weather, delay-free and personalized financial professional consultation services are provided for company staff, and the pressure of financial operation and maintenance staff is reduced to a certain extent, so that the enterprise operation and maintenance cost investment is reduced.

The foregoing is merely a preferred embodiment of the present invention, and it should be noted that it will be apparent to those skilled in the art that several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the scope of the invention.

Claims (10)

1. An intelligent dialogue method based on natural language processing comprises the following steps:

s1: semantic parsing and intent recognition;

judging the intention of a user according to the text input by the user by adopting a mode of combining a text retrieval mode, a natural sentence pattern matching mode, a rule sentence pattern matching mode and a meaning type matching shallow semantic understanding mode with deep learning mode, and understanding the requirement of the user; the system has a semantic association function, a user inputs fuzzy semantics, and the system automatically brings up maintained expansion questions to guide the user to accurately express intention;

s2: intent matching;

according to the S1 intention recognition result, similarity matching is carried out between the user problem and the similarity condition of the knowledge points in the system knowledge base in the high-dimensional semantic space by using sentence vectors;

s3: emotion analysis;

According to the S1 semantic analysis result, a user emotion is perceived in real time through a system built-in emotion model, and a corresponding answer is made according to the user emotion state;

s4: searching a knowledge base;

according to the results of S2 and S3, firstly extracting keywords, then searching a knowledge base by adopting a word vector model and a semantic similarity calculation method, outputting a text with highest similarity, and simultaneously, moisturizing the text by utilizing an emotion model;

s5: a multi-modal answer;

Generating an answer to the information matched with the S4 according to a predefined rule and a template, and replying by a text and document form aiming at the data query class; aiming at financial system and system operation consultation class, replying in the form of text, picture and file; meanwhile, the system has a question-associated function, and synchronously outputs the associated questions according to analysis of the user questions while outputting answers;

S6: modeling a customer portrait;

the system has the capability of counting, analyzing and classifying the problems, automatically generates customer portraits according to the consultation records of the users, and displays 'guessing you want to ask', 'hot spot problems', 'my portraits' on a dialogue interface, so that the users can more conveniently and quickly know the business services related to the customer portraits; according to the analysis result of the consultation frequency, the system automatically sorts the hot spot problems, and the user realizes quick consultation through the hot spot problem list.

S7: machine learning;

The system provides channels for demand collection and user feedback in the interactive window, the dialogue flow and the system setting function module; the system is internally provided with a machine learning platform, realizes algorithm and model level tuning, and continuously optimizes own answer and response capacity according to the machine learning algorithm.

2. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S1, the system is improved in accuracy of recognizing user intention by configuring threshold, weight and inheritance meaning parameters through a system built-in NLP engine.

3. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S1, the system uses a combination of shallow semantic understanding modes such as text retrieval, natural sentence pattern matching, regular sentence pattern matching and deep learning to calculate the similarity of the user problem and knowledge points in the knowledge base in the high-dimensional semantic space by using sentence vectors, so as to further judge the intention of the user.

4. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S1, a semantic association menu function is added on a dialogue interface, a user inputs fuzzy semantics, a system automatically brings up maintained expansion questions, guides the user to accurately express intention, accurately positions problems and provides correct answers, and the practicability of the system is improved.

5. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S1, the system is displayed on the user desktop in a three-dimensional robot image, and the user can enter the consultation dialogue interface by double-clicking the robot image, so that a convenient query mode is provided for the user.

6. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S2, the system has a menu authority control function, and only a user having menu authority can perform a corresponding operation; after S1 intention recognition, the system checks whether the user has the corresponding use authority of the consultation report, and if the authority check passes, the system replies correspondingly; if the permission verification is not passed, the system replies: to the contrary, you do not have the relevant use rights and ask the contact system personnel to apply for.

7. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S3, a system embeds a 30+ emotion model and supports a custom emotion model, a customer emotion is perceived in real time, and emotion pacifying is actively carried out in dialogue, so that customer experience is greatly improved.

8. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S4, a huge knowledge base is constructed by knowledge graph technology, including knowledge and information in financial field and office field; the knowledge base provides a plurality of modes for updating the domain resources and has an intention checking function, and if the following five conditions exist, the part of core words are not imported; comprising the following steps: the intended name repeats or exceeds the length; sentence pattern is not in accordance with rules or sentence pattern repetition; only importing corresponding intentions and expansion questions when the intended word slots do not exist; the action keywords and the business keywords are not intended to be imported when the action keywords and the business keywords are not present; the intent scene name already exists and the intent is automatically merged with the existing scene.

9. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S5, the system supports collecting a plurality of slot values, inheriting slot values and checking slot values in a plurality of rounds of dialogue, and has the capability of extracting and inheriting the slot values simultaneously in the same dialogue flow, so that the dialogue is more intelligent.

10. The intelligent dialogue method based on natural language processing according to claim 1, wherein in S5, the answer is generated according to the predefined rules and templates for the information matched with S4, and the content forms include text, document, table, map, link, and video.