WO2020114269A1 - 一种智能投顾的实现方法及系统 - Google Patents

一种智能投顾的实现方法及系统 Download PDF

Info

Publication number
WO2020114269A1
WO2020114269A1 PCT/CN2019/120573 CN2019120573W WO2020114269A1 WO 2020114269 A1 WO2020114269 A1 WO 2020114269A1 CN 2019120573 W CN2019120573 W CN 2019120573W WO 2020114269 A1 WO2020114269 A1 WO 2020114269A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
data
financial
information
financial database
Prior art date
Application number
PCT/CN2019/120573
Other languages
English (en)
French (fr)
Inventor
林震亚
杜新凯
刘勇
牛国扬
屠要峰
高洪
黄震江
Original Assignee
中兴通讯股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中兴通讯股份有限公司 filed Critical 中兴通讯股份有限公司
Publication of WO2020114269A1 publication Critical patent/WO2020114269A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/06—Asset management; Financial planning or analysis

Definitions

  • This disclosure relates to, but is not limited to, the field of communications.
  • Intelligent investment consulting refers to the network virtual artificial intelligence products based on customers' own financial needs, asset conditions, risk tolerance, risk preference and other factors, using modern portfolio theory, building data models through algorithms, using artificial intelligence technology and network platforms to provide financial management Consultancy services to replace traditional manual investment consultants.
  • intelligent investment advisory uses artificial intelligence to use portfolio theory, such as CAPM (Capital Assets Pricing Model), to formulate investment portfolios for users.
  • CAPM Capital Assets Pricing Model
  • the traditional intelligent investment advisor is to build a data model according to modern asset portfolio theory, and its asset allocation process is completely dependent on the Internet.
  • the accuracy of traditional intelligent investment advisory analysis is not high, and it is difficult to provide convenient and fast humanized services.
  • An aspect of an embodiment of the present disclosure provides an implementation method of intelligent investment consulting, including: processing data from multiple heterogeneous data sources based on Natural Language Processing (NLP), and adding the obtained data information to finance Database; determine the user's intention based on the information input by the user; and provide financial services to the user based on the user's intention based on the financial database.
  • NLP Natural Language Processing
  • an intelligent investment advisory system including: a data processing module, a deep semantic understanding module, a result feedback module, and a financial database
  • the data processing module is configured based on NLP for multiple heterogeneous data Process the source data and add the obtained data information to the financial database; and, obtain and process the information input by the user and send the processed information to the deep semantic understanding module
  • the deep semantic understanding module is configured to process the processed information To determine the user's intention
  • the result feedback module is configured to provide financial services to the user based on the user's intention based on the financial database.
  • a further aspect of an embodiment of the present disclosure provides an intelligent investment advisory system, including a memory, a processor, and a computer program stored on the memory and executable on the processor.
  • the processor implements the program to implement the aforementioned intelligent investment advisory implementation method.
  • Yet another aspect of the embodiments of the present disclosure provides a computer-readable storage medium on which one or more computer programs are stored, and one or more computer programs can be executed by one or more processors to implement the above-mentioned intelligent investment Gu's method of implementation.
  • FIG. 1 is an architectural diagram of an intelligent investment advisory system provided by an embodiment of the present disclosure.
  • FIG. 2 is a schematic diagram of the composition of a data processing module provided by an embodiment of the present disclosure.
  • FIG. 3 is a schematic diagram of text structuring processing provided by an embodiment of the present disclosure.
  • FIG. 4 is a schematic diagram of the composition of a deep semantic understanding module provided by an embodiment of the present disclosure.
  • FIG. 5 is an architectural diagram of a reading comprehension model provided by an embodiment of the present disclosure.
  • FIG. 6 is a schematic diagram of the composition of a result feedback module provided by an embodiment of the present disclosure.
  • FIG. 7 is a schematic diagram of a user portrait provided by an embodiment of the present disclosure.
  • FIG. 8 is a schematic diagram of a product portrait provided by an embodiment of the present disclosure.
  • FIG. 9 is an architecture diagram of a multi-modal joint learning model provided by an embodiment of the present disclosure.
  • FIG. 10 is a schematic diagram of an attention model provided by an embodiment of the present disclosure.
  • FIG. 11 is a schematic structural diagram of a meta-judgment network provided by an embodiment of the present disclosure.
  • FIG. 12 is a schematic diagram of the composition of a data model management module provided by an embodiment of the present disclosure.
  • FIG. 13 is a flowchart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 14 is a flowchart of processing data from multiple heterogeneous data sources according to an embodiment of the present disclosure.
  • 15 is a flowchart of joint learning provided by an embodiment of the present disclosure.
  • 16 is a flowchart of determining a user's intention provided by an embodiment of the present disclosure.
  • FIG. 17 is a flow chart of recommending products for users based on the knowledge base, product portraits, and user portraits in a financial database using meta-learning methods provided by an embodiment of the present disclosure.
  • FIG. 19 is another flowchart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 20 is another flowchart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 21 is another flow chart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 22 is still another flowchart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 23 is another flowchart of an implementation method of intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 24 is another flowchart of processing data from multiple heterogeneous data sources according to an embodiment of the present disclosure.
  • FIG. 25 is still another flowchart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 26 is another flow chart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • FIG. 27 is a schematic diagram of the composition of an intelligent investment advisory system provided by an embodiment of the present disclosure.
  • the embodiments of the present disclosure use NLP (Natural Language Processing) technology to extract and analyze events.
  • NLP Natural Language Processing
  • the embodiments of the present disclosure provide an interactive intelligent investment advisory system, which can complete the characterization of user attributes, and can analyze massive financial texts and data through NLP technology, and give appropriate responses according to user requirements.
  • the embodiments of the present disclosure can continuously improve and improve themselves based on meta-learning.
  • the intelligent investment consulting system may include: a data processing module 11, a deep semantic understanding module 12, a result feedback module 13 and a financial database 14.
  • the data processing module 11 may be configured to process data from multiple heterogeneous data sources based on NLP, and add the obtained data information to the financial database 14; and, obtain information input by the user for processing And send the processed information to the deep semantic understanding module 12.
  • the data processing module 11 can obtain data from multiple heterogeneous data sources, such as obtaining data from a stable financial service API (Application Programming Interface), or crawl all kinds of news on the Internet , Comments, etc. Then, through the text structured processing flow, useful information is stored in the financial database 14.
  • the text is mainly classified based on industry and genre, among which the genre includes news, comments, research reports, etc.
  • the financial database 14 may be a distributed database, which may include a structured database, a knowledge base, a graph library, a user portrait, a product portrait, etc.; where the graph library is a graph database, including a knowledge graph and multiple Affair map.
  • the knowledge graph is associated with the knowledge base.
  • FIGS. 2 and 3 it is a schematic diagram of the composition of the data processing module 11 provided by the embodiment of the present disclosure and a schematic view of the text structuring process provided by the embodiment of the present disclosure.
  • the data processing module 11 may include a data parser 111, a text analyzer 112, and an information extractor 113.
  • the data parser 111 may be configured to perform data parsing on data from multiple heterogeneous data sources to obtain text data.
  • data analysis may include: html (Hyper Text Markup Language, Hypertext Markup Language) analysis, word analysis, pdf analysis, etc.
  • the text analyzer 112 may be configured to perform text analysis on the text data to obtain text information.
  • text analysis may include: text cleaning, chapter structure analysis, syntax analysis, expression recognition, picture extraction, and the like.
  • chapter structure analysis can use deep learning classifiers, statistical language models, syntactic analysis can take the form of paragraph analysis and syntactic trees, expression recognition can carry out expression recognition and emotion analysis, picture extraction can carry out picture recognition and picture semantic understanding, etc. .
  • the level can be divided into three levels.
  • the first level refers to the internal structure, meaning and organizational form of the text, such as text analysis based on conversation analysis and corpus;
  • the second level refers to the modalities of the text embodied in language and other languages.
  • the third layer refers to the relationship between language and society, culture, and context, such as intermediary discourse analysis and genre analysis, critical discourse analysis, and communicative ethnology.
  • the dimensions can be divided into five dimensions.
  • the first is to put the center on the ontology of the discourse, and to study the characteristics and discourse structure of the discourse itself, such as conversation analysis and genre analysis.
  • the second is to focus on the content presented in the remaining articles, including social activities and social events and the role of building society, such as critical discourse analysis and positive discourse analysis.
  • the third is to focus on the individual's identity construction, including studying the individual's dialect, gender, class, and ethnic language.
  • the fourth is to focus on the relationship between text and text, that is, intertextuality research; some discuss the relationship between text and context, such as the theory of textual analysis of systemic functional linguistics.
  • the fifth is to explore text as a process and product in social practice, such as intermediary discourse analysis.
  • Classifiers are designed for different levels and dimensions, so that the structure of the chapter can be comprehensively analyzed.
  • Common text classifiers include FastText, TextCNN (text CNN), TextRNN (text RNN), TextRCNN (text RCNN), Hierarchical Attention Network (layered attention network), Seq2seq with Attention (sequence to sequence model with attention mechanism) ), Transformer (Transformer Network), Dynamic Memory (Network Dynamic Memory), Entity Network (EntityNet, Entity Network).
  • Considering model performance and model correlation usually using low correlation model integration can get better results.
  • the embodiments of the present disclosure can use the integration of FastText, TextCNN, EntityNet, DynamicMemory, and Transformer.
  • the model provided by the present disclosure can be managed in a unified manner, so it can be easily replaced.
  • the information extractor 113 may be configured to extract information based on text information and add the obtained structured text data to the structured database in the financial database 14.
  • information extraction based on text information may include: named entity recognition, relationship extraction, event extraction, table information extraction, etc.
  • named entity recognition may include entity disambiguation, LSTM+CRF (Long Short-Term Memory+Conditional Random Field, long-short-term memory network + conditional random field), rule extraction based on entity library, etc.
  • relationship/event extraction may include dynamic Convolutional neural networks, event recognition with attention (Attention), reinforcement learning, remote supervision with external knowledge added, etc.
  • table information extraction can include table structure analysis, table alignment, table completion, etc.
  • the traditional event extraction using dynamic convolutional network + attention mechanism method can already get good results.
  • the focus is on the collection and labeling of data sets, especially the reasonable use of bilingual corpus.
  • the embodiments of the present disclosure may use machine learning models (such as Transformer, Seq2seq with Attention) to process bilingual corpus, thereby expanding the corpus in disguise.
  • machine learning models such as Transformer, Seq2seq with Attention
  • the copy mechanism can be used to solve the overlap problem, so the main model becomes the encoder-decoder model with the copy mechanism.
  • a new element (Cell) needs to be used for encoding, but only one decoder is used to decode all triples during decoding.
  • both the extracted features in the text and the processed text will be stored in the structured database.
  • Text features (such as entities, relationships, events, etc.) stored in the database can be used to assist in the construction of the atlas.
  • the text analyzer 112 may be configured to perform text analysis on the information input by the user to obtain the text information input by the user.
  • the information input by the user may be voice, text, etc. If the information input by the user is voice, the voice may be converted into text.
  • the information extractor 113 may be configured to extract information according to text information input by the user, and acquire keywords and sentence patterns.
  • the data processing module 11 may further include a text summary submodule 114.
  • the text summary submodule 114 may be configured to perform summary processing on the text information obtained by the text analyzer 112, and then send it to the information extractor for information extraction.
  • the text summary sub-module 114 can change the long text into short text, so as to facilitate further analysis and event extraction.
  • the long text and the short text can be defined according to preset rules, for example, texts with more than 500 words are defined as long texts, and texts with less than 500 words are defined as short texts.
  • the text summary sub-module 114 may perform summary processing after the text analyzer 112 processes the text, or may perform summary processing after the data parser 111 processes the data, or may perform summary processing during the processing of the text analyzer 112.
  • the text summarization submodule 114 can be implemented by traditional summarization methods such as TextTeaser and Lexrank and the Seq2seq method based on the Copy mechanism.
  • text summary submodule There are two ways to call the text summary submodule. One is to call a specific module according to a specific problem. Traditional methods such as TextTeaser and Lexrank can control the reduction of the summary by modifying hyperparameters, and Seq2seq (Sequence to Sequence, sequence to sequence) The method needs to train multiple models to solve this problem; the second is to select a specific module for summarization according to the number of words in the text. For example, for the interface A, texts with a limit of more than 500 words call the Lexrank module uniformly.
  • the deep semantic understanding module 12 may be configured to determine the user's intention according to the processed information.
  • the deep semantic understanding module 12 may include an intent recognition sub-module 121.
  • the intent recognition sub-module 121 may be configured to determine an entity and an entity based on intent recognition rules and classification algorithms, combined with user portraits, according to keywords, sentence patterns, combined with entity extraction and text classification algorithms. intention.
  • the intent recognition sub-module 121 can also be configured to determine that the user is not performing investment-independent input (not chatting) based on the keywords and sentence patterns before determining the entity and intent according to the keywords and sentence patterns, and that the user is not in the business process in.
  • the deep semantic understanding module 12 may further include a sentence generation sub-module 122 that can be configured to determine that the user is performing an investment irrelevant to the investment based on the keywords and sentence patterns in the intent recognition sub-module 121 When inputting (in chatting), a chatting sentence or a paraphrase sentence is generated and fed back to the user through the result feedback module 13.
  • a sentence generation sub-module 122 that can be configured to determine that the user is performing an investment irrelevant to the investment based on the keywords and sentence patterns in the intent recognition sub-module 121 When inputting (in chatting), a chatting sentence or a paraphrase sentence is generated and fed back to the user through the result feedback module 13.
  • the sentence generation sub-module 122 can use the mainstream Seq2seq method to generate sentence, which can be used for chatting or retelling the sentence, increasing the intelligence of the system.
  • the intent recognition sub-module 121 determines that the user is performing input that is not related to the investment (in chatting) based on the keywords and sentence patterns, if it is determined that the user is asking for general questions such as date and weather, a general template can be used to pass the result feedback module 13 Feedback to users.
  • the deep semantic understanding module 12 further includes a reading comprehension sub-module 123.
  • the reading comprehension sub-module 123 may be configured to use data from the financial database 14 in the manner of reading comprehension for questions that occur many times and cannot be answered Get the corresponding answer in the message.
  • the reading comprehension sub-module 123 can find relevant documents based on the text and questions, through its entity and intent, and obtain corresponding answers by means of reading comprehension, which can be added to the knowledge base after manual review.
  • QA Question-Answer, Question-Answer
  • in the knowledge base can be added to the knowledge graph or affair graph in the graph database of the financial database 14 for common entities or events in the financial field.
  • the reading comprehension sub-module 123 can be implemented based on R-net, SLQA (Semantic Learning for Question Answering, based on layered fusion attention mechanism) and other models.
  • FIG. 5 it is an architectural diagram of a reading comprehension model provided by an embodiment of the present disclosure.
  • the reading comprehension model as an example based on the SLQA architecture, which has surpassed humans in specific tasks. Due to the complexity of the financial field, the system needs to pay attention to the extraction of relevant features when using this architecture. In theory, the finer the financial text features, the better the results. For events that occur multiple times, templates can be used to extract them, thereby increasing accuracy.
  • the model can be divided into coding layer, attention layer, matching layer and output layer.
  • Coding layer used for representation learning, it can be understood as a language model layer, which is used to transform texts and problems from discrete characters into semantic representation vectors. Multi-class deep learning methods are used for feature extraction.
  • Attention layer After obtaining effective question and chapter representations, locate the answer process based on the question, narrow the search range of alternative answers, restrict the search space through the attention mechanism, mainly perform multi-layer fusion attention expression, and focus on questions and chapters. Relevance alignment (Align), and constantly add global information (Fusion), each alignment is based on the underlying information and more refined on this basis, the method used is Co-Attention (chapter to problem, question to chapter) , Self-Attention (problem itself, chapter itself).
  • the attention mechanism is calculated and its weight and embedding result are fused (addition or splicing). For the fused result, multiply it with the embedding result for semantic representation, then put it together with the domain features of the text (vector representation) for self-arrangement (splicing), and then perform P2P (paragraph to paragraph) on the matrix ) Attention mechanism calculation.
  • the processing method is similar to the feature of the text, the difference is that no additional domain feature information needs to be introduced.
  • Matching layer used for problem and text matching after fusing information.
  • Bilinear matrix is used to learn the matching parameters of text and problem after multi-layer information filtering. Since the irrelevant information has been filtered in the previous stage, the final matching can be Complete the positioning of the answer. It should be noted that when processing text features, the results obtained by the P2P attention mechanism and the previous semantic representation should be weighted together, and this step is not required when processing the problem features.
  • Output layer Combine the matching information to mark the words in the chapter, and predict the probability that the corresponding word is the start or end position of the answer. After that, the model will extract a continuous text with the highest probability as the answer.
  • each part of the structure can be replaced with other modules that can achieve similar feature processing.
  • the coding layer can use multi-gram (N-gram) feature training language models
  • the attention layer can use multi-head attention (Multi-head Attention). structure.
  • the result feedback module 13 may be configured to provide the user with corresponding financial services based on the user's intention based on the financial database 14.
  • the user's intention may include the need to provide financial problem services, the need to provide data query services, the need for business processing services, and the need for investment advice services.
  • the feedback module 13 may include a response generation sub-module 131.
  • the response generation sub-module 131 may be configured as a deep semantic understanding module 12 to determine the user’s need to provide financial problem services according to the user’s intention, query the atlas library in the financial database 14, and determine that there is a corresponding The answer is output.
  • the reply generation sub-module 131 After clarifying the user's intention, the reply generation sub-module 131 obtains a reasonable reply through data retrieval, graph reasoning, or the results of various models.
  • the response generation sub-module 131 may be configured to: when querying the knowledge graph in the graph library and determining that there are matching entities and intentions, determine the corresponding by mapping relationship with the knowledge base in the financial database Answer, output the answer; or, query the knowledge graph in the graph library to determine that there are no matching entities and intents, query the affairs graph in the library, determine that there are corresponding events, perform inference analysis according to the affairs graph, and output the answer.
  • the result feedback module 13 may further include a similarity matching sub-module 132, which may be configured to conduct the information entered by the user and standard questions based on the knowledge base in the financial database 14 Similarity matching.
  • the similarity matching sub-module 132 can train the cold start similarity algorithm through the combination of traditional similarity features, and train the corresponding deep learning model according to the domain data to perform the similarity matching algorithm; the former has stronger domain mobility and robust stability The latter has higher accuracy in certain fields.
  • the similarity matching algorithm can search QA pairs in the knowledge base, and can also be used as a feature to assist other modules.
  • the response generation sub-module 131 can also be configured to: the deep semantic understanding module 12 determines that the user needs to provide financial question services according to the user’s intention, and matches the sub-module 132 according to the similarity when it is determined that there is no corresponding answer based on the graph library in the financial database 14 For the matching result, when the similarity is greater than or equal to the threshold, the answer corresponding to the standard question is output.
  • the result feedback module 13 further includes an interaction sub-module 133, which can be configured to determine whether the information entered by the user contains an entity or intent in the atlas library when the similarity is less than the threshold, based on Entity or intent to ask questions; or when it is determined that the user's question does not contain the entity or intent in the Atlas, follow the preset rules for general replies, replies, or recommendations.
  • an interaction sub-module 133 can be configured to determine whether the information entered by the user contains an entity or intent in the atlas library when the similarity is less than the threshold, based on Entity or intent to ask questions; or when it is determined that the user's question does not contain the entity or intent in the Atlas, follow the preset rules for general replies, replies, or recommendations.
  • the interaction sub-module 133 may be configured to actively guide the interaction to determine the user's intention when the field and intention of the user's information are unclear.
  • the module will memorize all the sentences of each round of dialogue and play the role of contextual interaction.
  • the reply generation sub-module 131 can also be configured as the deep semantic understanding module 12 to query the financial database 14 according to the keywords in the information input by the user when determining that the user needs to provide data query service according to the user's intention and output search result.
  • the response generation sub-module 131 can also be configured as the deep semantic understanding module 12 to determine the type of business that the user needs to handle when determining the user needs business processing service according to the user's intention At the time, perform business processing services.
  • the result feedback module may further include a product selection sub-module 134, which may be configured as a deep semantic understanding module 12 to determine the user’s need for investment advice services according to the user’s intention
  • the type of investment is based on the meta-learning method and recommends products for users based on the knowledge base, product portraits and user portraits in the financial database.
  • the product selection sub-module 134 can analyze various products and give a reasonable asset combination, while preventing excessive hot money from flowing into an industry.
  • the financial database 14 may include a graph library and a knowledge base.
  • the graph library may include a knowledge graph and a plurality of affair graphs, and the knowledge graph is associated with the knowledge database.
  • the intelligent investment advisory system may further include: a graph module 15 that may be configured to add entities or events in the knowledge base related to the financial field to the graph database in the financial database 14.
  • Graphs are mainly divided into financial knowledge graphs and affair graphs.
  • Knowledge graphs are mainly constructed based on financial professional knowledge, artificially constructed in the early stage, and later based on relationship extraction, entity recognition and intention recognition, based on a large number of text messages to improve, each industry, each category Financial products will serve as entities, each with its own attributes.
  • the affair graph constructs the important event flow on the basis of the knowledge graph, in the form of a directed and circular graph, nodes represent events, and directed edges represent succession and causality between events.
  • the entire graph library is composed of a knowledge graph and multiple affair graphs. Eventually, you can get the impact of different events on different industries or different indexes, and give expectations for upcoming events.
  • the knowledge graph can be logically divided into two layers: the model layer and the data layer.
  • the data layer is mainly composed of a series of facts, and knowledge will be stored in units of facts. Use (entity 1, relationship, entity 2), (entity, attribute, attribute value) to express facts, and choose graph database as storage medium, such as open source Neo4j, Twitter FlockDB, Sones GraphDB, etc.
  • the pattern layer is built on the data layer, mainly through the ontology library to regulate a series of fact expressions in the data layer.
  • Ontology is a conceptual template of structured knowledge base.
  • the knowledge base formed by ontology database not only has strong hierarchical structure, but also has less redundancy.
  • the initial financial knowledge graph adopts a top-down construction method, establishes a graph framework through the financial knowledge system and the experience of experts, and fills in the framework with universal entities, relationships and attributes.
  • the specific construction process is as follows: 1. Data integration of structured data and third-party databases, extraction of entities, relationships and attributes of semi-structured data and unstructured data; 2. Alignment of entities, relationships and attributes by knowledge reasoning ; 3. Use entities, relationships and attributes for ontology construction; 4. Perform quality assessment on the constructed ontology, and store it in the knowledge graph if passed; 5.
  • knowledge is updated, also adopt quality assessment, and pass the graph if passed Update.
  • the construction process of the affair graph is similar to the knowledge graph.
  • the affair graph defines the relationship between two kinds of events: one kind of inheritance and one kind of cause and effect. Both of these relations have a chronological order.
  • the research object of knowledge graph is nominal entity and its relationship
  • the research object of affairs graph is predicate event and its relationship.
  • the main knowledge forms of knowledge graphs are entity attributes and relationships
  • the affairs graphs are logical relations and probability transfer information.
  • the evolutionary relationship between events is mostly uncertain, while the relationship between entities is basically stable.
  • the intelligent investment advisory system may further include: a user portrait module 16, which may be configured to create or perfect a user portrait according to the user's attribute information.
  • the establishment of user portraits can help intent understanding of user sentences, and can also give better investment suggestions based on the user's risk preferences.
  • the user's attribute information may include at least one of the following: age factor, family income, investable amount, family burden, investment experience, acceptable loss, psychological factors, user target data, user interaction data, user behavior data, and interests.
  • product portraits can also be included in the financial database.
  • the establishment of product portraits can make it easier to evaluate products while giving users more accurate recommendations.
  • the analysis results of multi-modal and multi-task models can be combined with products Connect to determine its impact under the current situation.
  • FIG. 8 it is a schematic diagram of a product portrait provided by an embodiment of the present disclosure.
  • Product attributes are mainly divided into product types, yields, maximum drawdowns, related industries, volatility, American Depository Receipts (ADR), Relative Strength Index (RSI), asset allocation, and effective frontiers , Asset correlation, enterprise status, other quantitative factors, etc.
  • each user and product is characterized by vectors and stored in the financial database.
  • the product vector will keep changing with the market changes, and the user vector will be constantly updated according to user behavior.
  • the data information may include multi-modal data
  • the intelligent investment advisory system may further include: a joint learning module 17, which may be configured to obtain multi-modal data from the financial database 14, based on the multi-modal input
  • the joint learning model establishes the association relationship between multimodal data and industries, and is stored in the knowledge base in the financial database 14.
  • the joint learning module 17 is based on a multi-modal input Joint-Learning model, which can establish the relationship between multi-modal data and various industries, such as changes in supply and demand, changes in related product prices, and so on. For example, after the announcement of the US election results, the impact on the domestic real estate industry, specific labels can be reflected in the form of quantitative data such as the stock price change rate. If this part of the result appears multiple times, it can be added to the map after manual review.
  • FIG. 9 it is an architectural diagram of a multi-modal joint learning model provided by an embodiment of the present disclosure.
  • the main input of this model is structured text data, macro market data and picture data.
  • picture data can also be used for target recognition.
  • the mainstream organizational structure multilayers can be used) for abstract feature extraction; then these features are linearly spliced and imported into the fully connected layer. For pictures that are easier to understand semantically, you can directly input the semantic representation of the text as structured text instead of network input.
  • Structured text data is first processed in three ways. One is to extract traditional features, including simple features such as how many words the text contains and more complex features such as topic models.
  • the third is to summarize the text. This part is extracted by combining the traditional method and the Seq2seq method. The results obtained by these two types of methods are quite different, so the results of combining the two types of methods are meaningful. Subsequently, the sentence encoder (sentence encoder) method is used to encode the sentence, and then the residual module and the TCN are used to perform feature extraction (the specific number of layers depends on the situation). The reason for not using TCN in the keyword part is that this part does not consider the word order, but it is necessary to encode the position of each keyword in the document and add it to the word representation to improve the effect.
  • sentence encoder sentence encoder
  • Macro market data includes Dow Jones index, exchange rate, deposit interest rate, loan interest rate, etc.
  • the different combinations of such data represent that the global financial markets are in different situations, so it is of great significance to add such data.
  • the processing of the characteristics of this type of data is relatively simple, just need to normalize it and import it into the fully connected layer.
  • an abstract representation of the current financial market can be obtained.
  • Embedding, hierarchical embedding, or block embedding can be used during embedding, and then the results are imported into the fully connected layer. When it is not a vector but a matrix, it can also be further processed by CNN or other stronger feature extraction layers.
  • the intelligent investment advisory system may further include: a meta-learning module 18, which may be configured to build an attention model based on the meta-learning method, and optimize the joint learning model according to the attention model.
  • the meta-learning module 18 can be divided into two parts.
  • the first part is to design a general attention model based on the meta-learning idea and to improve the multi-task joint learning model of multi-modal input; the second part uses a meta-critic network (Meta-critic Network) )
  • the idea of learning is better loss (Loss).
  • This part mainly uses reinforcement learning to predict human behavior in different scenarios, thereby assisting system decision-making and improving the accuracy and intelligence of the system.
  • FIG. 10 it is a schematic diagram of an attention model provided by an embodiment of the present disclosure.
  • the basic idea of the first part is: people's attention can be improved by using previous experience, then use the previous tasks to train an Attention model, so that in the face of new tasks, you can directly focus on the most important parts.
  • This system constructs an Attention mechanism, and the final label judgment is obtained by superimposing Attention, and Attention is obtained by combining training after the transformation of historical samples and new sample functions.
  • the basic purpose is to use existing tasks to train a good Attention model.
  • This module can train a classification model separately and combine it with the results obtained by joint learning (Ensemble); it can also be embedded in a multi-modal joint learning model for joint training, thereby improving the final result.
  • This method is mainly used for categories with fewer training samples. When used, it is necessary to serially encode the historical samples of the category, and then extract the feature of the encoded information to enter the middle layer. For new samples, in addition to encoding and importing into the middle layer, Attention calculations (encoding methods such as Multi-head Attention) need to be performed with the encoding results of historical samples. The calculated weight is between the new samples and the historical samples. Degree of correlation, so that the impact of new samples on various industries can be judged from the impact of historical samples on various industries.
  • the second part is relatively independent.
  • the state space is established through the analysis of the current situation and the affair graph, and then the behavior space and feedback are established based on each state, and the best behavior of the person is finally obtained.
  • This part is mainly combined with the product selection sub-module 134 to better predict the financial market, because the behavior of most people will be reflected in the market in the future. It is assumed here that the holders of most funds are sensible, that is, the rich are smart enough.
  • FIG. 11 it is a schematic diagram of a meta-judgment network structure provided by an embodiment of the present disclosure. It can be seen from Figure 11 that the system constructs a meta-judgment network to learn to predict the loss of the Actor Network.
  • the meta-judgment network may include Meta-Value Network (MVN) and tasks- Behavior encoding (Task-Actor Encoder, TAEN).
  • meta-learning steps When using the meta-judgment network, it can be divided into meta-learning steps and meta-testing steps.
  • the meta-learning steps can be shown in Table 1.
  • Table 1 The steps of meta-learning.
  • Table 2 The steps of the meta test.
  • the intelligent investment advisory system may further include: a data model management module 19.
  • the data model management module 19 may include a dialog management submodule 191 and a model management submodule 192.
  • the dialog management sub-module 191 may be configured to store dialog data, including original text data and processed feature data, for effective session management.
  • the model management sub-module 192 may be configured to maintain an algorithm library, a rule database and a financial database required by the data processing module, deep semantic understanding module, joint learning module and meta-learning module.
  • the intelligent investment advisory system has the functions of financial market analysis and intelligent chat robots.
  • the embodiments of the present disclosure establish a connection with various industries through a multi-modal joint learning model, so that changes in the entire financial market can be analyzed through recent news and data.
  • the model uses offline training, considering that network feed-forward is time-consuming.
  • the system analyzes the text and data of the day at regular intervals, and saves important results for easy recall.
  • the meta-learning module 18 has two main functions. The first is to improve the model to solve the small sample problem; the second is to learn human behavior through the results of the previous model.
  • the intelligent chat robot part of the system can better understand the user's intention through data processing and semantic understanding combined with the user's portrait.
  • the user's portrait can be continuously improved and revised.
  • this system is more intelligent and user-friendly.
  • embodiments of the present disclosure also provide a method for implementing intelligent investment advisory.
  • FIG. 13 it is a flowchart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • the method may include: step 201 and step 202.
  • step 201 the data of the multiple heterogeneous data sources is processed based on natural language processing NLP, and the obtained data information is added to the financial database.
  • step 202 the user's intention is determined according to the information input by the user, and based on the financial database, the financial service is provided to the user according to the user's intention.
  • step 201 since step 201 may be executed periodically or in real time, the execution order of steps 201 and 202 is not limited.
  • processing data from multiple heterogeneous data sources based on NLP may include steps 301-step 303.
  • step 301 data is analyzed to obtain text data.
  • step 302 perform text analysis on the text data to obtain text information.
  • step 303 information extraction is performed based on the text information to obtain structured text data.
  • the data from multiple heterogeneous data sources can be all kinds of news, reviews, research reports, macro market data, etc.
  • data parsing may include at least one of the following operations: html parsing, word parsing, and pdf parsing.
  • text analysis includes at least one of the following operations: text cleaning, chapter structure analysis, syntax analysis, expression recognition, and picture extraction.
  • chapter structure analysis can use deep learning classifiers, statistical language models, syntactic analysis can take the form of paragraph analysis and syntactic trees, expression recognition can carry out expression recognition and emotion analysis, picture extraction can carry out picture recognition and picture semantic understanding, etc. .
  • information extraction based on text information includes at least one of the following operations: named entity recognition, relationship extraction, event extraction, and table information extraction.
  • entity recognition it can include entity disambiguation, LSTM+CRF, entity library-based rule extraction, etc.
  • relationship/event extraction can include dynamic convolutional neural networks, attentional event recognition, reinforcement learning, and the addition of external knowledge. Remote supervision, etc.
  • table information extraction can include table structure analysis, table alignment, table completion, etc.
  • the method before performing information extraction based on text information, the method further includes: performing digest processing on the text information.
  • Summary processing refers to changing long text into short text, which is convenient for further analysis and event extraction.
  • the long text and the short text can be defined according to preset rules, for example, texts with more than 500 words are defined as long texts, and texts with less than 500 words are defined as short texts.
  • Abstract processing is performed after text analysis, or after data analysis, or during text analysis.
  • the data information may include multi-modal data
  • the method may further include: obtaining multi-modal data from a financial database, establishing a multi-modal data and industry based on a joint learning model of multi-modal input The relationship between them is stored in the knowledge base in the financial database.
  • the joint learning model based on multimodal input can establish the relationship between multimodal data and various industries, such as changes in supply and demand, changes in related product prices, and so on. For example, after the announcement of the US election results, the impact on the domestic real estate industry, specific labels can be reflected in the form of quantitative data such as the stock price change rate. If this part of the result appears multiple times, it can be added to the map after manual review.
  • the multimodal data includes structured text data, macro market data, and picture data.
  • Acquiring multi-modal data from the financial database, and establishing an association relationship between the multi-modal data and the industry based on the multi-modal input joint learning model may include steps 401-404.
  • step 401 structured text data, macro market data, and picture data are obtained from the financial database.
  • step 402 feature extraction of multimodal data is performed.
  • step 403 the features of the multimodal data are embedded, and the embedded features are fully connected.
  • step 404 the fully connected result is embedded with the strong correlation feature corresponding to each industry, and joint optimization is performed to establish the association relationship between the multimodal data and the industry.
  • the method may further include: building an attention model based on meta-learning, and optimizing the joint learning model according to the attention model.
  • an Attention mechanism is constructed, and the final label judgment is obtained by superposition of Attention, and Attention is obtained by combined training after the transformation of historical samples and new sample functions.
  • the basic purpose is to use existing tasks to train a good Attention model.
  • This module can train a classification model separately and combine it with the results obtained by joint learning; it can also be embedded in a multi-modal joint learning model for joint training, thereby improving the final result.
  • This method is mainly used for categories with fewer training samples. When used, it is necessary to serially encode the historical samples of the category, and then extract the feature of the encoded information to enter the middle layer. For new samples, in addition to encoding and importing into the middle layer, Attention calculations (encoding methods such as Multi-head Attention) need to be performed with the encoding results of historical samples. The calculated weight is between the new samples and the historical samples. Degree of correlation, so that the impact of new samples on various industries can be judged from the impact of historical samples on various industries.
  • the financial database may include a graph library and a knowledge base
  • the graph database may include a knowledge graph and multiple affair graphs, and the knowledge graph is associated with the knowledge database.
  • the method may further include: adding entities or events in the knowledge base related to the financial field to the atlas library in the financial database.
  • the initial knowledge graph adopts a top-down construction method, establishes a graph framework through the financial knowledge system and the experience of experts, and fills in the framework with universal entities, relationships and attributes. Later, as knowledge is continuously updated, people's cognitive abilities are continuously improved, and the graph will be updated and iterated in a bottom-up manner.
  • the financial database may include a user portrait
  • the method may further include: establishing or perfecting the user portrait based on the user's attribute information; wherein, the user's attribute information includes at least one of the following: age factor, family income, Investable amount, family burden, investment experience, acceptable loss, psychological factors, user target data, user interaction data, user behavior data, interests and hobbies.
  • the establishment of user portraits can help intent understanding of user sentences, and can also give better investment suggestions based on the user's risk preferences.
  • determining the user's intention according to the information input by the user may include step 501 and step 502.
  • step 501 according to the information input by the user, keywords and sentence patterns are acquired.
  • step 502 based on the intent recognition rules and classification algorithm, combined with the user portrait, the entities and intents are determined according to keywords and sentence patterns.
  • the method may further include: determining that the user is not performing and investing based on the keywords and sentence patterns Irrelevant input (not chatting), and it is determined that the user is not in the business process.
  • a general template can be used to feed back the response to the user.
  • the method may further include: acquiring a corresponding answer from the data information of the financial database by reading and comprehending the question that appears many times and cannot be answered.
  • the method may further include: adding the audited questions and the corresponding answers to the knowledge base in the financial database.
  • QA Quality-Answer in the knowledge base can be added to the knowledge graph or the affairs graph in the graph database of the financial database if it involves common entities or events in the financial field.
  • providing the user with corresponding financial services according to the user's intention may include: when determining that the user needs to provide financial problem services according to the user's intention, query the atlas database in the financial database to determine the existence The corresponding answer is output.
  • the user's intention may include the need to provide financial problem services, the need to provide data query services, the need for business processing services, and the need for investment advice services.
  • outputting the answer may include: querying the atlas of knowledge in the atlas database to determine whether there are matching entities and intentions through The mapping relationship in the knowledge base in the database determines the corresponding answer and outputs the answer; or, when querying the knowledge graph in the atlas to determine that there are no matching entities and intents, query the atlas in the atlas to determine the corresponding event , Infer analysis according to the affair graph, and output the answer.
  • providing the user with corresponding financial services according to the user's intention may include: determining that the user needs to provide financial problem services according to the user's intention, and determining that there is no corresponding based on the graph library in the financial database
  • the information input by the user is matched with the similarity of the standard question, and when the similarity is greater than or equal to the threshold, the answer corresponding to the standard question is output.
  • the cold start similarity algorithm is trained through the combination of traditional similarity features, and the similarity matching algorithm is trained based on the domain data to train the corresponding deep learning model; the former has stronger domain mobility and robust stability, and the latter Higher accuracy in specific fields.
  • the similarity of the information input by the user is matched with the standard question.
  • the method may further include: when it is determined that the information input by the user includes an entity or intent in the atlas library, Ask questions based on entities or intentions; or, when it is determined that the user's questions do not include entities or intentions in the Atlas, follow the preset rules for general replies, replies, or recommendations.
  • providing corresponding financial services to the user according to the user's intention may include: when determining that the user needs to provide a data query service according to the user's intention, query according to the keyword in the information entered by the user Financial database, output query results.
  • the method may further include: when determining the entity or intent in the atlas library that contains the financial database in the information entered by the user, asking questions based on the entity or intent; Or, when it is determined that the user's question does not include the entity or intent in the atlas library, the general replies, replies, or recommendation information are conducted according to the preset rules.
  • providing the user with corresponding financial services according to the user's intention may include: determining the type of business the user needs to handle according to the user's intention When it is determined to support the handling of the business, the business handling service shall be executed.
  • providing the user with corresponding financial services according to the user's intention may include: determining the type of investment required by the user when determining the user's investment recommendation service according to the user's intention;
  • the way of learning is based on the knowledge base, product portraits and user portraits in the financial database to recommend products for users.
  • FIG. 17 it is a flow chart of recommending products for users based on the knowledge base, product portraits, and user portraits in the financial database provided by the embodiment of the present disclosure using meta-learning.
  • using meta-learning to recommend products for users based on the knowledge base, product portraits, and user portraits in the financial database may include steps 601-605.
  • step 601 based on the knowledge base, product portraits, and user portraits in the financial database, an environment, behavior, and state space are constructed.
  • step 602 an optimization goal is constructed.
  • step 603 task-behavior coding is performed to calculate the rewards brought by different behaviors in different states.
  • step 604 the task-behavior coding is embedded in the meta-value network to learn the loss function of the task.
  • step 605 strategy-gradient training is performed to optimize the behavior in a specific environment and a specific state, and recommend products for users based on the optimized behavior.
  • FIG. 18 it is a flowchart of deep semantic understanding and result feedback provided by an embodiment of the present disclosure.
  • the process of deep semantic understanding and result feedback may include steps 701-728.
  • step 701 it is determined whether the user is chatting; if it is, the information entered by the user is not related to finance, then it is transferred to the chatting part for processing, and step 702 is performed; if not, step 705 is performed.
  • step 702 it is determined whether the user asks general questions such as date and weather; if so, step 703 is performed; if not, step 704 is performed.
  • step 703 the general template is used to feed back the reply to the user.
  • step 704 by means of sentence generation, a chat statement or a paraphrase sentence is generated, and the reply is fed back to the user.
  • step 705 if it is not in a chat, priority is given to whether it is still in the process; if it is currently in a business process, step 706 is performed; if it is confirmed that it is not in the business process, then step 708 is performed.
  • step 706 it is determined whether to terminate the process; if so, step 708 is performed; if not, step 707 is performed.
  • step 707 the user is guided to complete the business process.
  • steps 706 to 707 if it is currently in a business process, the user is first guided to complete the business process unless the user wants to actively terminate the process.
  • step 708 the user's intention is identified based on the user's portrait and context.
  • step 709 it is determined whether the intention is clear. If it is not clear, step 710 is performed; if it is clear, different processing procedures are adopted for different types of problems; when determining financial problems, step 711 is performed; when determining that it is data query, step 719 is performed ; When judging to conduct business, execute step 722; When judging to make investment advice, execute step 726.
  • step 710 a general rhetorical question is asked to be accurate.
  • step 711 search the knowledge graph to determine whether there are corresponding entities and intentions; if so, find the corresponding answer through the mapping relationship with the knowledge base, and perform step 718; if not, perform step 712.
  • step 712 the affair graph is searched, if there is a corresponding event, inference analysis is performed based on the affair graph, and step 718 is performed; if not, step 713 is performed.
  • step 713 similarity matching is performed based on the knowledge base and the standard question.
  • step 714 it is determined whether there is a standard problem greater than the threshold, and if so, step 718 is performed; if not, step 715 is performed.
  • step 715 the map is searched for whether it contains the corresponding entity or intention. If so, step 716 is performed; if not, step 717 is performed.
  • step 716 a rhetorical question is asked based on the entity or intent in the map.
  • step 717 a general rhetorical question is asked to tell the user that there is no answer for the question at the same time, at the same time, some hot questions with high similarity can be recommended to the user, and the interaction continues.
  • step 718 the user is responded accordingly.
  • step 719 the query content is determined through keyword extraction.
  • step 720 it is judged that the financial database releases the content containing the query, if step 721 is performed; if not, step 715 is performed.
  • step 721 the user is given the corresponding result. Users can obtain related content by clicking the interface.
  • step 722 the type of business that the user needs to handle is determined.
  • step 723 it is determined whether the business process is supported, and if so, step 724 is performed; if not, step 725 is performed.
  • step 724 enter the business process.
  • step 725 the user is fed back a reply that cannot be processed.
  • step 726 the type of investment required by the user is determined.
  • step 727 the meta-learning module and the product portrait are combined to determine which products the most recent form is beneficial to.
  • step 728 the user's investment proposal is combined with the user's portrait.
  • the call of the long text algorithm is not involved in this process, because the long text processing time is too long, which will cause the user to wait, thereby affecting the experience.
  • the long text processing is done offline, and the analysis results are stored in the financial database, which is convenient for the above procedure.
  • Example 1 Bank intelligent customer service system
  • the system can be used in bank virtual customer service. Different from the traditional intelligent customer service system, this system can better answer financial-related questions or give investment advice. For different users, it can be combined with multi-dimensional portraits of users to give a better understanding of intent. When targeting the banking business area, relevant content should be added to the knowledge base and atlas. As shown in FIG. 19, it is another flowchart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure. The method may include steps 801-806.
  • step 801 based on the user's personal information, historical deposits and loans, and purchase of wealth management products, combined with preset questions, a user profile is initially established.
  • step 802 it is determined whether the user is chatting, if yes, enter the chatting module; if not, it is determined whether the business process is in progress. If the user says: "I want to apply for a credit card", it is judged that it is not chat.
  • step 803 if the business process is in progress, the user is guided to complete the business process; if not, the user's intention is understood through the user portrait. For the previous problem, it will automatically turn to intent recognition.
  • step 804 intent recognition is performed based on the user's portrait and context. If the intent is clear, the process is continued; if it is not clear, a rhetorical question is asked. For "I want to apply for a credit card”, the intention is clear, and the system recognizes it as a business process, then the business process begins.
  • step 805 the service type to be handled by the user is identified according to methods such as semantic template and similarity calculation.
  • the business identified as "credit card processing” is supported by the bank.
  • step 806 interact with the user according to the process, and guide the user to handle the business. All interaction records will be saved, which is convenient for the semantic understanding of the following text and the improvement of user portraits.
  • Example 2 Provide product information query service
  • the system can be used to provide product information query services. As shown in FIG. 20, it is another flow chart of a method for implementing intelligent advisory provided by an embodiment of the present disclosure.
  • the method may include steps 901-905.
  • step 901 a user portrait is created according to a preset question to perform intent recognition.
  • the first few steps are similar to application example 1. If the user asks: "I want to view ZTE's stocks and related research reports", this problem is identified as a product information query problem, so the data query process is followed.
  • step 902 the keywords in the user's question are mined, and the entity and intention are determined by the intention classification method.
  • the entities can be identified as “stocks” and “research reports”, the intent is “inquiry”, and the limited scope is "ZTE".
  • step 903 the system receives the message. If it cannot find it, it asks the question based on the entity or intent; if it can be found, it returns a link or button, and after clicking it, you can enter the graphical interface.
  • the computer side is similar to the wind (financial data and analysis tool service provider) form, and the mobile terminal is similar to the stock display interface of the straight flush.
  • the entire interface is as concise as possible, providing only the information the user wants. For example, only the stock trend and related research report list are provided in this question. Of course, the user can obtain other information through further clicking or searching in the interface.
  • step 904 if the user closes the interface, the user is deemed to terminate the process.
  • the user's behavior record will be recorded, which can be used to improve the user's portrait and facilitate the semantic understanding of the following text.
  • step 905 information that appears multiple times in the problem and cannot be queried in the system will be recorded in the background, and a new data source can be added after manual review.
  • Example 3 Provide financial market analysis services
  • the system can be used to provide financial market analysis services. As shown in FIG. 21, it is another flow chart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • the method may include steps 1001 to 1007.
  • step 1001 a user portrait is established according to a preset question to perform intention recognition.
  • the first few steps are similar to application example 1. If the user asks, "what impact will the tsunami sweep across Shanghai”? This problem is identified as a financial-related problem, so the financial problem process is followed.
  • step 1002 the keywords in the user's question are mined, and the entity and intention are determined by the intention classification method.
  • the entities are “tsunami” and “Shanghai”, and the intention is "impact”.
  • step 1003 search the knowledge graph to see if the corresponding entity intent combination can be found, and if so, return the answer; if not, go to the next link.
  • the above questions do not belong to traditional financial knowledge points, and do not correspond to specific industries or products, so the knowledge map cannot be answered.
  • step 1004 the affair map is searched to confirm whether the corresponding event can be found.
  • the above problem belongs to the incident of "natural disaster in Shanghai". If it has been stored in the atlas, it can return to the series of consequences that will occur after the incident. It is assumed here that the event has not been stored in the affair graph, and then enters the similarity calculation link.
  • step 1005 similarity calculation is performed with the standard problem in the knowledge base. If the standard question is matched, the corresponding result is returned based on the QA pair; if not, the rhetorical question is entered.
  • the similarity calculation is mainly divided into two categories.
  • the traditional feature combination method can be used for cold start.
  • the DSSM Deep Structured Semantic Model, Deep Structure Semantic Model
  • step 1006 it is determined whether the user question includes an entity or an intention existing in the knowledge graph; if it exists, a rhetorical question is asked based on the entity or intention; if it does not exist, a general rhetorical question is asked.
  • a rhetorical question is asked based on the entity or intention; if it does not exist, a general rhetorical question is asked.
  • the queried question is related to the recent event as much as possible. The event is extracted based on Figure 3. And through Figure 9 to establish its connections with various industries.
  • step 1007 the questions that appear many times and cannot be answered temporarily by the system will be recorded in the background, and the corresponding answers will be given offline by way of FIG. 5. The higher quality of such replies will be judged by manual review whether they are added to the graph or the knowledge base.
  • the system can be used to provide investment advice. As shown in FIG. 22, it is another flow chart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure. The method may include steps 1101 to 1105.
  • step 1101 a user portrait is created according to a preset question to perform intention recognition.
  • the first few steps are similar to application example 1. If the user asks: "I want to buy a financial product, what is the right option?" This problem is identified as the need to provide investment advice, so the investment advice process is followed.
  • step 1102 the keywords in the user's question are mined, and the entity and intention are determined using the intention classification method.
  • the entity is "financial product” and the intention is "purchase”, it is determined that the user wants to purchase a financial product.
  • step 1103 based on the model in Figure 9, determine which industries are beneficial to recent text news and macro data (this part of the results will be calculated and stored offline), and find relevant financial products based on the product library and evaluate Its relevance.
  • step 1104 corresponding product recommendations are given based on the user portrait. If the user averses risk, he mainly recommends low-risk and low-yield products.
  • users can make independent combinations based on the recommended products, and user behaviors will be recorded in the background.
  • the user can also give feedback on the recommendation results (star rating) to further improve the user portrait.
  • FIG. 23 it is another flow chart of a method for implementing intelligent advisory provided by an embodiment of the present disclosure.
  • the method may include steps 1201 to 1205.
  • step 1201 the interface will provide options for querying the database, knowledge graph, and affair graph, which the user can click to enter. It can also be called through an external interface.
  • the database contains structured text data and macro market data.
  • the former is the result after processing according to Figure 3, and the latter includes various financial product trends and macro indexes.
  • step 1203 all kinds of financial knowledge points are stored in the knowledge graph.
  • connections between various industries will also be recorded in the graph.
  • step 1204 the affair graph will record the typical financial behavior process, mainly including some periodic financial events. This part will continue to improve with the self-learning of the system.
  • step 1205 the user's calling process to the database will also be recorded in the background to further improve the user's portrait.
  • FIG. 24 another flowchart of processing data from multiple heterogeneous data sources provided by an embodiment of the present disclosure may include steps 1301-step 1309.
  • step 1301 the file obtained from the data source is first parsed.
  • the html file uses an html parser
  • the pdf file uses a pdf parser.
  • the parser extracts useful text and pictures in the file.
  • step 1302 enter the text analyzer, first perform text cleaning, and then perform coarse-grained text classification through the chapter analysis method.
  • step 1303 the text is syntactically analyzed to extract relevant paragraphs and syntactic features.
  • step 1304 the expressions in the text are recognized and converted into text, and text sentiment analysis is performed based on this.
  • step 1305 the pictures in the file are extracted and analyzed using the semantic understanding of the pictures (RCNN+RNN).
  • step 1306 enter the information extractor, first use rules and deep learning methods to identify named entities of the text, and extract the corresponding entities.
  • step 1307 the relationship and event extraction of the text mainly uses dynamic convolutional networks, reinforcement learning and other methods.
  • step 1308 structural analysis is performed on the parsed table file, and then complete table information is extracted through table alignment and completion.
  • step 1309 the cleaned text data and the extracted features and information are classified and stored in a structured database.
  • FIG. 25 it is another flow chart of a method for implementing intelligent investment advisory provided by an embodiment of the present disclosure.
  • the method may include steps 1401 to 1405.
  • step 1401 structured text data, macro market data, and picture data are obtained from the database, respectively, and data sets with different time spans are selected according to the type of problem to be processed. For example: “How will the demand for the steel industry change in the near future?", you only need to read the recent data.
  • step 1402 feature extraction of multi-modal data is performed, and different feature extraction methods are adopted for different types of data.
  • the text needs to be summarized, and then the sentence vector of the summarized text is subjected to feature extraction.
  • step 1403 the features of the multimodal data are embedded. There are multiple ways, such as stitching, weighted sum after normalization, attention, etc. Then fully connect the embedded features.
  • step 1404 the results of the previous layer are processed and embedded into the strong correlation features corresponding to each industry.
  • the strong correlation features here can be extracted for a specific industry using the method of FIG. 3, and the types are not limited. Such as: the financial report data of listed companies in the steel industry, the events extracted from the steel sector research report, the emotions of the people in the steel sector in the stock bar, etc. Different types of data have their corresponding feature extraction methods, which are similar to 1402.
  • step 1405 the result obtained in 1404 is processed, and after activation, it is imported into the loss function layer, and the optimization target is a combination of all loss functions.
  • each industry has different numbers and types of indicators, and they are classified according to time span, such as long-term, medium-term, and short-term, and different joint learning models are trained according to the indicators of different spans.
  • FIG. 26 it is another flowchart of an implementation method of intelligent advisory provided by an embodiment of the present disclosure.
  • the method may include steps 1501 to 1505.
  • step 1501 environment, behavior, and state space are constructed.
  • the current situation is favorable to certain indicators in certain industries, such as the current situation, various product combinations recommended to users.
  • the status refers to which products the user has held and how many current assets they have.
  • step 1502 an optimization goal is constructed. Such as the user's annualized expected income.
  • an optimization goal is constructed.
  • step 1503 task-behavior coding is performed to calculate the rewards brought by different behaviors in different states.
  • step 1504 the task-behavior coding is embedded in the meta-value network, and the loss function of the entire task is learned.
  • step 1505 strategy-gradient training is performed to optimize the behavior in a specific environment and state.
  • an embodiment of the present disclosure also provides an intelligent investment advisory system.
  • the system may include a memory 1601, a processor 1602, and a computer program 1603 stored in the memory 1601 and running on the processor 1602.
  • the processor When 1602 executes the program, it implements the intelligent investment advisory implementation method provided by the present disclosure.
  • Embodiments of the present disclosure also provide a computer-readable storage medium that stores computer-executable instructions, and the computer-executable instructions are used to perform an intelligent investment advisory implementation method provided by the implementation of the present disclosure.
  • the above storage medium may include, but is not limited to: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc.
  • ROM read-only memory
  • RAM random access memory
  • mobile hard disk magnetic disk or optical disk, etc.
  • computer storage media includes both volatile and nonvolatile implemented in any method or technology for storing information such as computer readable instructions, data structures, program modules, or other data Sex, removable and non-removable media.
  • Computer storage media include but are not limited to RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cartridges, magnetic tape, magnetic disk storage or other magnetic storage devices, or may Any other medium for storing desired information and accessible by a computer.
  • the communication medium generally contains computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transmission mechanism, and may include any information delivery medium .

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Operations Research (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • Technology Law (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Business, Economics & Management (AREA)
  • Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种智能投顾的实现方法及系统，该方法包括：基于自然语言处理对来自多元异构数据源的数据进行处理，并将得到的数据信息加入金融数据库（201）；根据用户输入的信息，确定用户的意图；以及基于金融数据库，根据用户的意图为用户提供金融服务（202）。

Description

一种智能投顾的实现方法及系统 技术领域
本公开涉及但不限于通信领域。
背景技术
智能投顾是指网络虚拟人工智能产品基于客户自身的理财需求、资产状况、风险承受能力、风险偏好等因素，运用现代投资组合理论，通过算法搭建数据模型，利用人工智能技术和网络平台提供理财顾问服务，取代传统的人工投资顾问。简单地说，智能投顾就是通过人工智能，使用投资组合理论，如CAPM(Capital Asset Pricing Model，资本资产定价模型)，来为用户制定投资组合。在互联网金融日益深入的今天，智能投顾有可能成为影响“资金-资产”配给的关键环节。
从原理上看，传统智能投顾是根据现代资产组合理论构建数据模型，其资产配置的过程完全依靠互联网来完成。但是，传统智能投顾分析准确性不高，也很难提供方便快捷的人性化服务。
发明内容
本公开实施例的一个方面提供一种智能投顾的实现方法，包括：基于自然语言处理(Natural Language Processing，NLP)对来自多元异构数据源的数据进行处理，并将得到的数据信息加入金融数据库；根据用户输入的信息，确定用户的意图；以及基于金融数据库，根据用户的意图为用户提供金融服务。
本公开实施例的另一方面提供一种智能投顾系统，包括：数据处理模块、深层语意理解模块、结果反馈模块和金融数据库，其中：数据处理模块，配置为基于NLP对来自多元异构数据源的数据进行处理，并将得到的数据信息加入金融数据库；以及，获 取并处理用户输入的信息，将处理后的信息发送至深层语意理解模块；深层语意理解模块，配置为根据处理后的信息，确定用户的意图；以及，结果反馈模块，配置为基于金融数据库，根据用户的意图为用户提供金融服务。
本公开实施例的再一方面提供一种智能投顾系统，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，处理器执行程序时实现上述的智能投顾的实现方法。
本公开实施例的又一方面提供一种计算机可读存储介质，其上存储有一个或者多个计算机程序，一个或者多个计算机程序可被一个或者多个处理器执行，以实现上述的智能投顾的实现方法。
附图说明
图1为本公开实施例提供的智能投顾系统的架构图。
图2为本公开实施例提供的数据处理模块的组成示意图。
图3为本公开实施例提供的文本结构化处理的示意图。
图4为本公开实施例提供的深层语意理解模块的组成示意图。
图5为本公开实施例提供的阅读理解模型的架构图。
图6为本公开实施例提供的结果反馈模块的组成示意图。
图7为本公开实施例提供的用户画像的示意图。
图8为本公开实施例提供的产品画像的示意图。
图9为本公开实施例提供的多模态联合学习模型的架构图。
图10为本公开实施例提供的注意力模型的示意图。
图11为本公开实施例提供的元-评判网络的结构示意图。
图12为本公开实施例提供的数据模型管理模块的组成示意图。
图13为本公开实施例提供的智能投顾的实现方法的一种流程图。
图14为本公开实施例提供的对来自多元异构数据源的数据进行处理的一种流程图。
图15为本公开实施例提供的联合学习的流程图。
图16为本公开实施例提供的确定用户的意图的流程图。
图17为本公开实施例提供的采用元学习的方式，基于金融数据库中的知识库、产品画像和用户画像，为用户进行产品推荐的流程图。
图18为本公开实施例提供的深层语义理解及结果反馈的流程图。
图19为本公开实施例提供的智能投顾的实现方法的另一种流程图。
图20为本公开实施例提供的智能投顾的实现方法的又一种流程图。
图21为本公开实施例提供的智能投顾的实现方法的再一种流程图。
图22为本公开实施例提供的智能投顾的实现方法的再一种流程图。
图23为本公开实施例提供的智能投顾的实现方法的再一种流程图。
图24为本公开实施例提供的对来自多元异构数据源的数据进行处理的另一种流程图。
图25为本公开实施例提供的智能投顾的实现方法的再一种流程图。
图26为本公开实施例提供的智能投顾的实现方法的再一种流程图。
图27为本公开实施例提供的智能投顾系统的组成示意图。
具体实施方式
下文中将结合附图对本公开的实施例进行详细说明。
在附图的流程图示出的步骤可以在诸如一组计算机可执行指令的计算机系统中执行。并且，虽然在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于此处的顺序执行所示出或描述的步骤。
由于金融市场与各类新闻，研报具备强关联性，故而本公开实施例采用NLP(Natural Language Processing，自然语言处理)技术对其进行事件抽取及分析。本公开实施例提供一种可交互式的智能投顾系统，可以完成用户属性刻画，并能通过NLP技术对海量金融文本及数据进行分析，根据用户的要求给出合适的回复。此外，本公开实施例可基于元学习不断进行自我提升和改进。
如图1所示，其为本公开实施例提供的智能投顾系统的架构图，该智能投顾系统可包括：数据处理模块11、深层语意理解模块12、结果反馈模块13和金融数据库14。
根据本公开提供的实施例，数据处理模块11，可配置为基于NLP对来自多元异构数据源的数据进行处理，并将得到的数据信息加入金融数据库14；以及，获取用户输入的信息进行处理，并将处理后的信息发送至深层语意理解模块12。
根据本公开提供的实施例，数据处理模块11可以从多元异构数据源获取数据，例如从稳定的金融服务API(Application Programming Interface，应用程序编程接口)获取数据，或者在网 上爬取各类新闻、评论等。然后通过文本结构化处理流程，将有用的信息存入金融数据库14中。文本主要基于行业及类型进行分类，其中，类型包括新闻、评论、研报等。
根据本公开提供的实施例，金融数据库14可以是分布式数据库，可以包括结构化数据库、知识库、图谱库、用户画像、产品画像等；其中，图谱库为图数据库，包括一个知识图谱和多个事理图谱。知识图谱与知识库相关联。
根据本公开提供的实施例，如图2和图3所示，其分别为本公开实施例提供的数据处理模块11的组成示意图和本公开实施例提供的文本结构化处理的示意图。数据处理模块11可包括数据解析器111，文本分析器112和信息抽取器113。
根据本公开提供的实施例，数据解析器111可配置为对来自多元异构数据源的数据进行数据解析，得到文本数据。其中，数据解析可以包括：html(Hyper Text Markup Language，超文本标记语言)解析、word解析、pdf解析等。
根据本公开提供的实施例，文本分析器112可配置为对文本数据进行文本分析，得到文本信息。
根据本公开提供的实施例，文本分析可以包括：文本清洗、章节结构分析、句法分析、表情识别、图片提取等。
其中，章节结构分析可采用深度学习分类器、统计语言模型，句法分析可采用段落分析和句法树的形式，表情识别可分别进行表情识别和情绪分析，图片提取可进行图片识别和图片语意理解等。
对于章节结构分析，可以从多层次和多维度进行分类。层次可主要分为三层，第一层指篇章内部的结构、意义和组织形式，如以会话分析和语料库为基础的篇章分析；第二层指语言和其他 语言体现的语篇的模态，如多模态话语分析；第三层指语言与社会、文化、语境的关系，如中介话语分析和体裁分析、批评话语分析和交际民族学等。
维度可主要分五维，第一种是把中心放在语篇本体上，研究的是语篇本身的特征及语篇结构，如会话分析和体裁分析。第二种是把重点放在余篇所呈现的内容上，包括社会活动和社会事件以及对于社会的构建作用上，如批评话语分析、积极话语分析等。第三种是把中心放在个体的身份构建上，包括研究个体的方言、性别、阶层、种族用语上。第四种是把重点放在语篇与篇之间的关系上，即互文性研究；有的探讨语篇、语境的关系等，如系统功能语言学的语篇分析理论。第五种是探讨语篇作为社会实践中的过程和产品，如中介话语分析。
针对不同层次和维度，分别设计分类器，从而可以全面的对篇章结构进行分析。常用的文本分类器有FastText、TextCNN(文本CNN)、TextRNN(文本RNN)、TextRCNN(文本RCNN)、Hierarchical Attention Network(分层注意力网络)、Seq2seq with Attention(带注意力机制的序列到序列模型)、Transformer(变形网络)、Dynamic Memory Network(DynamicMemory，动态记忆网络)、Entity Network(EntityNet，实体网络)。考虑到模型性能和模型相关性，通常采用相关性低的模型集成可以得到更好的结果，本公开实施例可采用FastText、TextCNN、EntityNet、DynamicMemory和Transformer的集成。当然，如果出现了更好的文本分类器，由于根据本公开提供的实施例可对模型进行统一管理，故而可轻易替换。
根据本公开提供的实施例，信息抽取器113可配置为根据文本信息进行信息抽取，并将得到的结构化文本数据加入金融数据库14中的结构化数据库。
其中，根据文本信息进行信息抽取，可包括：命名实体识别、关系抽取、事件抽取、表格信息抽取等。
其中，命名实体识别可包括实体消歧、LSTM+CRF(Long Short-Term Memory+Conditional Random Field，长短期记忆网络+条件随机场)、基于实体库的规则抽取等；关系/事件抽取可包括动态卷积神经网络、带注意力(Attention)的事件识别、强化学习、加入外部知识的远程监督等；表格信息抽取可包括表格结构分析、表格对齐、表格补全等。
对于关系抽取，传统事件抽取采用动态卷积网络+注意力机制的方法已经可以得到不错的效果，提升重点主要在数据集的搜集和标注层面，特别是在双语语料的合理利用。本公开实施例可以采用机器学习模型(如Transformer、Seq2seq with Attention)对双语语料进行处理，从而变相扩充了语料。当句子中出现多个实体和多个关系的情况时，可以给句子里的每个词打上语义标签，找三元组，于是问题转换为打标签任务。在这种情况下，可采用复制(Copy)机制解决重叠(Overlap)问题，所以主模型变为了加入Copy机制的编码器-解码器(Encoder-Decoder)模型。该模型中，对于每个三元组，需要采用新的元素(Cell)进行编码，但解码时只采用一个解码器解码所有三元组。
根据本公开提供的实施例，文本中提取的特征与处理后的文本都将存在结构化数据库中。数据库中保存的文本特征(如实体、关系、事件等)可用于辅助图谱的构建。
根据本公开提供的实施例，针对用户输入的信息，文本分析器112可配置为对用户输入的信息进行文本分析，得到用户输入文本信息。
其中，用户输入的信息可以是语音、文字等内容，如果用户输入的信息是语音，则可将该语音转换为文字。
根据本公开提供的实施例，信息抽取器113可配置为根据用户输入文本信息进行信息抽取，获取关键词和句式。
根据本公开提供的实施例，数据处理模块11还可包括文本摘要子模块114。
文本摘要子模块114可配置为将文本分析器112得到的文本信息进行摘要处理，再发送至信息抽取器进行信息抽取。
文本摘要子模块114可将长文本变为短文本，从而便于进一步分析和事件抽取。其中，长文本和短文本可以根据预设的规则定义，例如，500字以上的文本定义为长文本，小于500字的文本定义为短文本。文本摘要子模块114可以在文本分析器112处理文本之后进行摘要处理，也可以在数据解析器111处理数据之后进行摘要处理，也可以在文本分析器112处理过程中进行摘要处理。
文本摘要子模块114可以通过TextTeaser、Lexrank这类传统摘要方法及基于Copy机制的Seq2seq方法实现。调用文本摘要子模块有两种方式，其一是根据特定问题指定调用特定模块，TextTeaser、Lexrank这类传统方法可通过修改超参数控制摘要的精简程度，而Seq2seq(Sequence to Sequence，序列到序列)方法需要训练多个模型来解决这个问题；其二是根据文本字数选择特定模块进行摘要，如对于接口A，限定500字以上的文本统一调用Lexrank模块。
根据本公开提供的实施例，深层语意理解模块12，可配置为根据处理后的信息，确定用户的意图。
如图4所示，其为本公开实施例提供的深层语意理解模块12的组成示意图。深层语意理解模块12可包括意图识别子模块121，意图识别子模块121可配置为基于意图识别规则和分类算法，结合用户画像，按照关键词、句式，结合实体抽取及文本分类算法 确定实体和意图。
另外，意图识别子模块121还可配置为按照关键词和句式确定实体和意图之前，基于关键词和句式确定用户不在执行与投资无关的输入(不在闲聊)，且确定用户不在业务办理流程中。
根据本公开提供的实施例，深层语意理解模块12还可包括语句生成子模块122，语句生成子模块122可配置为在意图识别子模块121基于关键词和句式确定用户在执行与投资无关的输入(在闲聊)时，生成闲聊语句或复述语句，并通过结果反馈模块13反馈至用户。
语句生成子模块122可采用主流Seq2seq方法进行语句生成，这部分可用于闲聊或是语句复述，增加系统的智能性。
如果意图识别子模块121基于关键词和句式确定用户在执行与投资无关的输入(在闲聊)时，若确定用户在询问日期、天气等通用问题，则可以采用通用模板，通过结果反馈模块13反馈至用户。
根据本公开提供的实施例，深层语意理解模块12还包括阅读理解子模块123，阅读理解子模块123可配置为对于多次出现且无法回答的问题，采用阅读理解的方式从金融数据库14的数据信息中获取相应的回答。
阅读理解子模块123可以根据文本及问题，通过其实体和意图找出相关文档，并采用阅读理解的方式获取相应回答，经人工审核后可加入知识库中。知识库中的QA(Question-Answer,问题-回答)对涉及金融领域常见实体或事件，可加入金融数据库14的图谱库中的知识图谱或事理图谱。
阅读理解子模块123可基于R-net、SLQA(Semantic Learning for Question Answering，基于分层融合注意力机制)这类模型实现。
如图5所示，其为本公开实施例提供的阅读理解模型的架构图。以阅读理解模型基于SLQA架构为例，该架构在特定任务上已经超越了人类。由于金融领域的复杂性，本系统在使用该架构时需要注意相关特征的提取，理论上金融文本特征做得越细，结果会越好。对于多次出现的事件，可采用模板形式提取，从而增加准确性。
该模型可主要分为编码层、注意力层、匹配层和输出层。
编码层：用于表示学习，可以理解为语言模型层，用以将篇章及问题从离散字符转变为蕴含语义的表征向量，采用了多类深度学习方法进行特征提取。
对于文本文件，首先需要训练词向量，优先采用词矩阵表征和预训练的语言模型(基于通用语料)解决一词多义的问题；然后分别采用词向量和字的独热码(One-Hot)表征训练当前文本的语言模型，得到其相关特征。此外，由于基于词、字嵌入(Embedding)学到的特征无法跨文本，故而还需要通过tf-idf和TexTrank提取关键词，并采用CNN对其进行处理。随后，将上述特征进行排列，并采用文本组织(Inception)结构对其进行进一步处理。
对于问题文本，同样需要采用词(词向量得到的方法同上)、字嵌入训练语言模型。由于问题通常较短，故而不需要关键词(每个词都很关键)，但可采用TCN(Temporal Convolutional Network，时间卷积网络)结构对其进一步进行特征抽取(其特征与语言模型区别较大)。随后，同样将上述特征进行排列，并采用文本组织结构对其进行进一步处理。
注意力层：得到有效的问题及篇章表征后，为表达依据问题定位答案过程，缩小备选答案查找范围，将搜索空间通过注意力机制约束，主要进行多层融合注意力表示，对问题和篇章进行相关性对齐(Align)，并不断补充全局信息(Fusion)，每一次对 齐都基于下层信息并在此基础上更加细化，采用的方式分别为Co-Attention(篇章到问题，问题到篇章)，Self-Attention(问题自身，篇章自身)。
在实际实现时，对于文本特征，首先将之前传入的特征进行嵌入(可采用全连接结构)，然后将问题的嵌入特征一起进行排列，并对排列后的结果进行Q2P(问题到段落)的注意力机制计算，并将其权重和嵌入结果进行融合(相加或者拼接)。对于融合后的结果，将其与嵌入结果相乘进行语义表示，然后将其与该文本的领域特征(向量表示)放在一起进行自排列(拼接)，然后对于该矩阵进行P2P(段落到段落)注意力机制计算。
对于问题特征，其处理方式与文本特征类似，区别是不需要引入额外的领域特征信息。
匹配层：用于做融合信息后的问题和篇章匹配，采用双线性矩阵来学习经过多层信息过滤后的篇章和问题匹配参数，由于在前一阶段无关信息已经被过滤，最后的匹配可完成答案的定位工作。需要注意的是，文本特征处理时，需将P2P注意力机制得到的结果与之前的语义表示一起进行权重融合，在处理问题特征时不需要这一步操作。
输出层：结合匹配信息对篇章中词汇进行标注，预测相应词汇是答案开始位置或结束位置的概率。之后，模型会抽取可能性最高的一段连续文本作为答案。
该结构中的每一部分都可替换为能实现类似特征处理的其它模块，如编码层可采用多元语法(N-gram)特征训练语言模型，注意力层可以采用多头注意力(Multi-head Attention)结构。
根据本公开提供的实施例，结果反馈模块13，可配置为基于金融数据库14，按照用户的意图为用户提供相应的金融服务。
用户的意图可以包括需要提供金融问题服务、需要提供数据查询服务、需要业务办理服务、需要投资建议服务等。
如图6所示，其为本公开实施例提供的结果反馈模块13的组成示意图。结果反馈模块13可包括回复生成子模块131，回复生成子模块131可配置为深层语意理解模块12按照用户的意图确定用户需要提供金融问题服务时，查询金融数据库14中的图谱库，确定存在相应的答案时，输出该答案。
在明确用户意图后，回复生成子模块131通过数据检索、图谱推理或是各类模型的结果得到合理的回复。
根据本公开提供的实施例，回复生成子模块131可配置为：查询图谱库中的知识图谱，确定有相匹配的实体和意图时，通过与金融数据库中的知识库中的映射关系确定相应的答案，输出答案；或者，查询图谱库中的知识图谱，确定没有相匹配的实体和意图时，查询图谱库中的事理图谱，确定有相应的事件，按照事理图谱进行推断分析，输出答案。
根据本公开提供的实施例，结果反馈模块13，还可包括相似度匹配子模块132，相似度匹配子模块132可配置为基于金融数据库14中的知识库，将用户输入的信息与标准问题进行相似度匹配。
相似度匹配子模块132可以分别通过传统相似度特征的组合训练冷启动相似度算法，以及根据领域数据训练相应的深度学习模型进行相似度匹配算法；前者具备更强的领域迁移性和鲁棒稳定性，后者在特定领域内拥有更高的精度。相似度匹配算法可以进行知识库中QA对的检索，同时也可以作为特征辅助其它模块。
回复生成子模块131还可配置为：深层语意理解模块12按照用户的意图确定用户需要提供金融问题服务，基于金融数据库14中的图谱库确定没有相应的答案时，根据相似度匹配子模块132的匹配结果，在相似度大于或等于阈值时，输出标准问题对应的 答案。
根据本公开提供的实施例，结果反馈模块13还包括交互子模块133，交互子模块133可配置为在相似度小于阈值时，确定用户输入的信息中包含图谱库中的实体或意图时，基于实体或意图进行反问；或者确定用户的问题中不包含图谱库中的实体或意图时，按照预设规则进行通用反问、回复或推荐信息。
交互子模块133可配置为当用户信息所在领域和意图不清楚时，主动引导交互确定用户意图。此外，该模块会记忆每轮对话的所有语句，起到上下文交互的作用。
根据本公开提供的实施例，回复生成子模块131还可配置为深层语意理解模块12按照用户的意图确定用户需要提供数据查询服务时，按照用户输入的信息中的关键词查询金融数据库14，输出查询结果。
根据本公开提供的实施例，回复生成子模块131还可配置为深层语意理解模块12按照用户的意图确定用户需要业务办理服务时，确定用户需要办理的业务类型，根据业务类型在确定支持办理业务时，执行业务办理服务。
根据本公开提供的实施例，结果反馈模块包括还可包括产品选择子模块134，产品选择子模块134可配置为深层语意理解模块12按照用户的意图确定用户需要投资建议服务时，确定用户需要的投资类型，采用元学习的方式，基于金融数据库中的知识库、产品画像和用户画像，为用户进行产品推荐。
产品选择子模块134可以对各类产品进行分析并给出合理的资产组合，同时防止某个行业流入过多热钱。
根据本公开提供的实施例，金融数据库14可包括图谱库和知识库，图谱库可包含一个知识图谱和多个事理图谱，知识图谱与 知识库相关联。
根据本公开提供的实施例，智能投顾系统还可包括：图谱模块15，可配置为将知识库中涉及金融领域的实体或事件加入金融数据库14中的图谱库中。
图谱主要分为金融知识图谱和事理图谱，知识图谱主要基于金融专业知识进行构建，前期人工构建，后期主要通过关系抽取、实体识别及意图识别，基于大量文本消息进行完善，每个行业、每类金融产品将会作为实体，分别带有各自的属性。事理图谱在知识图谱基础上对于重要事件流进行构建，形式为有向有环图，结点代表事件，有向边代表事件之间的顺承、因果关系。整个图谱库由一个知识图谱和多个事理图谱组成，最终可以得到不同事件对不同行业或是不同指数分别会产生怎样的影响，且会对即将发生的事件给出预期。
知识图谱在逻辑上可分为模式层与数据层两个层次，数据层主要是由一系列的事实组成，而知识将以事实为单位进行存储。用(实体1，关系，实体2)、(实体、属性，属性值)这样的三元组来表达事实，选择图数据库作为存储介质，例如开源的Neo4j、Twitter的FlockDB、Sones的GraphDB等。模式层构建在数据层之上，主要是通过本体库来规范数据层的一系列事实表达。本体是结构化知识库的概念模板，通过本体库而形成的知识库不仅层次结构较强，并且冗余程度较小。
需要注意的是，初期金融知识图谱采用自顶向下的构建方式，通过金融知识体系及专家的经验建立图谱框架，并在框架内填入普适性的实体、关系及属性。后期，随着知识不断更新，人的认知能力不断提升，图谱将采用自底向上的方式进行更新迭代。具体构建过程如下：1、对结构化数据及第三方数据库进行数据整合，对半结构化数据及非结构化数据进行实体、关系及属性抽取；2、 通过知识推理对实体、关系及属性进行对齐；3、采用实体、关系及属性进行本体构建；4、对构建好的本体进行质量评估，如通过便存入知识图谱中；5、当知识更新时，同样采取质量评估，如通过便进行图谱更新。
事理图谱构建流程与知识图谱类似，所不同的是事理图谱定义了两种事件间关系：一种顺承，一种因果，这两种关系都有时间顺序。此外，知识图谱研究对象为名词性实体及其关系，事理图谱研究对象是谓词性事件及其关系。知识图谱主要知识形式是实体属性和关系，事理图谱则是事理逻辑关系以及概率转移信息。事件间的演化关系多数是不确定的，而实体之间的关系基本是稳定的。
根据本公开提供的实施例，智能投顾系统还可包括：用户画像模块16，可配置为根据用户的属性信息建立或完善用户画像。
用户画像的建立可有助于对用户语句进行意图理解，还可以根据用户的风险偏好给出更好的投资建议。
如图7所示，其为本公开实施例提供的用户画像的示意图。用户的属性信息可包括如下至少之一：年龄因素、家庭收入、可投资金额、家庭负担、投资经验、可接受亏损、心理因素、用户目标数据、用户交互数据、用户行为数据、兴趣爱好。
另外，类似用户画像，金融数据库中还可包括产品画像，产品画像的建立可以在更方便对产品进行评估的同时给用户更精确的推荐，同时可以将多模态多任务模型的分析结果与产品关联，判断其在当前形势下受到的影响。
如图8所示，其为本公开实施例提供的产品画像的示意图。产品属性主要分为产品类型、收益率、最大回撤、关联产业、波动率、美国存托凭证(American Depository Receipt，ADR)、相对强弱指数(Relative Strength Index，RSI)、资产配置、有效前 沿、资产相关度、企业状况、其它量化因子等。
得到上述指标后，将其向量化，即每个用户及产品由向量进行表征，并存入金融数据库中。产品向量将随市场的变化而不断变化，而用户向量将根据用户行为不断更新。
根据本公开提供的实施例，数据信息可包括多模态数据，智能投顾系统还可包括：联合学习模块17，可配置为从金融数据库14中获取多模态数据，基于多模态输入的联合学习模型，建立多模态数据与行业之间的关联关系，存入金融数据库14中的知识库中。
联合学习模块17基于多模态输入的联合学习(Joint-Learning)模型，可建立多模态数据与各个行业之间的联系，如供需变化、相关产品价格变化等等。如：美国大选结果公布后，对国内房地产行业的影响，具体标签可通过板块股价等量化数据变化率的形式体现。该部分结果若多次出现，经人工审核后可加入图谱。
如图9所示，其为本公开实施例提供的多模态联合学习模型的架构图。本模型的主要输入为结构化的文本数据、宏观市场数据和图片数据，图片数据除了可采用R-CNN(Region-CNN)系列或是YOLO(You Only Look Once)方法进行目标识别，还可采用主流的组织结构(可采用多层)进行抽象特征提取；然后将这些特征线性拼接后导入全连接层。对于较容易进行语义理解的图片，可不作为网络输入，直接将文本语义表示作为结构化文本输入。结构化文本数据先采用三种方法进行处理，其一是提取传统特征，包括文本包含多少词这类简单特征以及主题模型这类较为复杂的特征，这部分提取的特征都是由数字或者向量进行表示，然后将这些特征进行归一化后拼接并导入全连接层(fc)。其二是采用tf-idf(Term Frequency–Inverse Document Frequency，词频-逆文本频率指数)、TextRank等方法进行文档关键词提取，并采用 Word2vec(Word to Vector，词向量)方法进行表示，通常采用FastText训练Skip-Gram模型可以得到最好的表征效果，然后将拼接后的矩阵导入多层ResNet模块进行特征抽取；这里采用ResNet的原因是关键词拼接得到的矩阵不用考虑前后相关性，而实际实验中发现该类结构能比TextCNN起到更好的特征提取效果。其三是对文本进行摘要，这部分采用传统方法和Seq2seq方法结合的方式提取，这两类方法得到的结果具有较大差异性，故而组合两类方法的结果具有意义。随后，采用句编码器(sentence encoder)的方法进行句编码，再分别采用残差模块和TCN进行特征抽取(具体层数视情况而定)。关键词部分不采用TCN的原因是这部分不考虑词序，但需要对每个关键词在文档中出现的位置进行编码，加入到词表征中，以提升效果。
宏观市场数据包括道琼斯指数、汇率、存款利率、贷款利率等。该类数据的不同组合代表着全球金融市场处于不同的形势下，故而加入这类数据有重要的意义。这类数据的特征处理较为简单，只需要将其归一化后导入全连接层即可。通过将这类数据和上述文本数据的嵌入，可得到当前金融市场的抽象表征，嵌入时可采用拼接嵌入、分层嵌入或者分块嵌入的方式，然后将其结果导入全连接层，当嵌入结果不是向量而是矩阵时，也可采用CNN或者其它更强的特征提取层对其进一步处理。
由于金融市场具有完整性，故而在本公开提供的实施例中将不同行业的输出模块放在一起训练，这就是典型的联合学习(Joint Learning)结构。只要不同任务中具有相关性，该结构已经证明了能取得比单模型更好的效果。在联合学习时，加入各个行业相关的强关联特征，从而可以更好的预测某天的文本数据在当前形势下会对某个行业带来何种影响。
根据本公开提供的实施例，智能投顾系统还可包括：元学习 模块18，可配置为基于元学习的方式建立注意力模型，根据注意力模型优化联合学习模型。
元学习模块18可以分两部分，第一部分基于元学习思想设计通用的注意力(Attention)模型，改进多模态输入的多任务联合学习模型；第二部分采用元-评判网络(Meta-critic Network)的思路学习更好的损失(Loss)，该部分主要通过强化学习在不同场景下进行人的行为预测，从而辅助系统决策，提升系统的准确性和智能性。
如图10所示，其为本公开实施例提供的注意力模型的示意图。第一部分的基本思路是：人的注意力是可以利用以往的经验来实现提升的，那么利用以往的任务来训练一个Attention模型，从而面对新的任务，能够直接关注最重要的部分。本系统构造一个Attention机制，最后的标签判断通过Attention的叠加得到，而Attention则通过历史样本及新样本函数变换后的组合训练得到。基本目的就是利用已有任务训练出一个好的Attention模型。该模块可以单独训练一个分类模型，与联合学习得到的结果进行组合(Ensemble)；也可以嵌入到多模态联合学习的模型中联合训练，从而提升最终结果。
该方法主要用于训练样本较少的类别，使用时，需要将该类别的历史样本进行序列化编码，然后对其编码信息进行特征提取，进入中间层。对于新样本，除了进行编码后导入中间层，还需与历史样本的编码结果进行Attention计算(可采用Multi-head Attention这类新型Attention方法)，计算得到的权重即为新样本与历史样本之间的关联度，从而可以从历史样本对各个行业的影响判断新样本对各个行业的影响。
第二部分较为独立，通过当前形势分析及事理图谱建立状态空间，再基于每个状态建立行为空间及反馈，最终得到人的最佳 行为。这部分主要是与产品选择子模块134结合，对金融市场进行更好的预判，因为大多数人的行为将在未来体现在市场上。这里假设大部分资金的持有者是理智的，即有钱人足够聪明。如图11所示，其为本公开实施例提供的元-评判网络结构示意图。从图11中可以看出本系统构造了一个元-评判网络来学习预测行为网络(Actor Network)的损失，该元-评判网络可包括元-价值网络(Meta-Value Network，MVN)和任务-行为编码(Task-Actor Encoder，TAEN)。
使用元-评判网络时，可分为元学习步骤及元测试步骤，元学习的步骤可如表1所示。
表1：元学习的步骤。
输入：任务生成器T；
输出：训练好的任务及价值网络；
1、初始化：任务和价值网络；
2、For episode＝1 to max episode do；
3、从T中生成任务M；
4、初始化M的策略网络(执行器)；
5、For step＝1 to max steps do；
6、采用小批量对任务进行采样；
7、For每一个小批量中的任务do；
8、从任务中采样训练数据；
9、训练特定任务执行器；
10、End；
11、训练价值网络；
12、训练任务网络；
13、End；
14、End。
元测试的步骤可如表2所示。
表2：元测试的步骤。
输入：没见过的任务，训练好的任务和价值网络；
输出：训练好的策略网络；
1、初始化：策略网络(执行器)；
2、For step＝1 to max step do；
3、从任务中采样训练数据；
4、训练执行器；
5、End。
根据本公开提供的实施例，智能投顾系统还可包括：数据模型管理模块19。
如图12所示，其本公开实施例提供的数据模型管理模19块组成示意图。数据模型管理模块19可包括对话管理子模块191和模型管理子模块192。
对话管理子模块191可配置为存储对话数据，包括原始文本数据及处理后的特征数据，进行有效的会话管理。
模型管理子模块192可配置为维护数据处理模块、深层语义理解模块、联合学习模块及元学习模块需要的算法库、规则库以及金融数据库等。
综上所述，本公开实施例提供的智能投顾系统，具有金融市场分析及智能聊天机器人的功能，通过从多元异构数据源获取数据，进行文本结构化处理，将有用的信息存入分布式数据库中。基于大量文本及各类宏观市场数据，本公开实施例通过多模态联 合学习模型建立其与各行业的联系，从而可以通过近期新闻和数据分析整个金融市场的变化。该模型采用离线训练，考虑网络前馈较为耗时，系统每隔一段时间对当天文本及数据进行分析，保存重要结果便于调用。元学习模块18主要有两个作用，第一个是改进模型，解决小样本问题；第二个是通过之前模型的结果学习人的行为。
图谱与语义理解相辅相成，图谱的知识点和事件可用于更好的对文本进行理解，而从文本中抽取出的新常见实体和关系可用于知识图谱的扩充。
系统中智能聊天机器人部分通过数据处理和语义理解，结合用户画像更好的理解用户意图。此外，通过用户在系统上的行为，可对用户画像进行不断的完善及修正。
与业界相关方案相比，本系统更加智能化、人性化。
参照前述针对智能投顾的描述，本公开实施例还提供一种智能投顾的实现方法。如图13所示，其为本公开实施例提供的智能投顾的实现方法的一种流程图，该方法可包括：步骤201和步骤202。
在步骤201中，基于自然语言处理NLP对多元异构数据源的数据进行处理，将得到的数据信息加入金融数据库。
在步骤202中，根据用户输入的信息，确定用户的意图，基于金融数据库，按照用户的意图为用户提供金融服务。
根据本公开提供的实施例，由于步骤201可以是定期或实时执行，所以步骤201和202的执行顺序不限。
根据本公开提供的实施例，通过基于NLP对来自多元异构数据源的数据进行处理，可以获取实时完善的金融数据信息，结合意图识别，可以为用户提供准确的金融市场分析及数据分析服务， 从而给用户更精准且方便快捷的金融服务。
如图14所示，其为本公开实施例提供的对来自多元异构数据源的数据进行处理的一种流程图。根据本公开提供的实施例，基于NLP对来自多元异构数据源的数据进行处理，可包括步骤301-步骤303。
在步骤301中，对数据进行数据解析，得到文本数据。
在步骤302中，对文本数据进行文本分析，得到文本信息。
在步骤303中，，根据文本信息进行信息抽取，得到结构化文本数据。
其中，来自多元异构数据源的数据可以是各类新闻、评论、研报、宏观市场数据等。
根据本公开提供的实施例，数据解析，可包括如下操作中的至少之一：html解析、word解析、pdf解析。
根据本公开提供的实施例，文本分析，包括如下操作中的至少之一：文本清洗、章节结构分析、句法分析、表情识别、图片提取。
其中，章节结构分析可采用深度学习分类器、统计语言模型，句法分析可采用段落分析和句法树的形式，表情识别可分别进行表情识别和情绪分析，图片提取可进行图片识别和图片语意理解等。
根据本公开提供的实施例，根据文本信息进行信息抽取，包括如下操作中的至少之一：命名实体识别、关系抽取、事件抽取、表格信息抽取。
其中，对于命名实体识别，可包括实体消歧、LSTM+CRF、基于实体库的规则抽取等；关系/事件抽取可包括动态卷积神经网络、带注意的事件识别、强化学习、加入外部知识的远程监督等； 表格信息抽取可包括表格结构分析、表格对齐、表格补全等。
根据本公开提供的实施例，根据文本信息进行信息抽取之前，还包括：对文本信息进行摘要处理。
摘要处理是指将长文本变为短文本，从而便于进一步分析和事件抽取。其中，长文本和短文本可以根据预设的规则定义，例如，500字以上的文本定义为长文本，小于500字的文本定义为短文本。在文本分析之后进行摘要处理，也可以在数据解析之后进行摘要处理，也可以在文本分析过程中进行摘要处理。
根据本公开提供的实施例，数据信息可包括多模态数据，该方法还可包括：从金融数据库中获取多模态数据，基于多模态输入的联合学习模型，建立多模态数据与行业之间的关联关系，存入金融数据库中的知识库中。
基于多模态输入的联合学习模型，可建立多模态数据与各个行业之间的联系，如供需变化、相关产品价格变化等等。如：美国大选结果公布后，对国内房地产行业的影响，具体标签可通过板块股价等量化数据变化率的形式体现。该部分结果若多次出现，经人工审核后可加入图谱。
如图15所示，其为本公开实施例提供的联合学习的流程图。根据本公开提供的实施例，多模态数据包括结构化文本数据、宏观市场数据和图片数据。从金融数据库中获取多模态数据，基于多模态输入的联合学习模型，建立多模态数据与行业之间的关联关系，可包括步骤401-步骤404。
在步骤401中，从金融数据库中获取结构化文本数据、宏观市场数据及图片数据。
在步骤402中，进行多模态数据的特征提取。
在步骤403中，将多模态数据的特征嵌入，将嵌入后的特征 进行全连接。
在步骤404中，将全连接的结果与每个行业对应的强关联特征嵌入，进行联合优化，建立多模态数据与行业之间的关联关系。
根据本公开提供的实施例，该方法还可包括：基于元学习的方式建立注意力模型，根据注意力模型优化联合学习模型。
参照图10，构造一个Attention机制，最后的标签判断通过Attention的叠加得到，而Attention则通过历史样本及新样本函数变换后的组合训练得到。基本目的就是利用已有任务训练出一个好的Attention模型。该模块可以单独训练一个分类模型，与联合学习得到的结果进行组合；也可以嵌入到多模态联合学习的模型中联合训练，从而提升最终结果。
该方法主要用于训练样本较少的类别，使用时，需要将该类别的历史样本进行序列化编码，然后对其编码信息进行特征提取，进入中间层。对于新样本，除了进行编码后导入中间层，还需与历史样本的编码结果进行Attention计算(可采用Multi-head Attention这类新型Attention方法)，计算得到的权重即为新样本与历史样本之间的关联度，从而可以从历史样本对各个行业的影响判断新样本对各个行业的影响。
根据本公开提供的实施例，金融数据库可包括图谱库和知识库，图谱库可包含一个知识图谱和多个事理图谱，知识图谱与知识库相关联。
根据本公开提供的实施例，该方法还可包括：将知识库中涉及金融领域的实体或事件加入金融数据库中的图谱库中。
初期知识图谱采用自顶向下的构建方式，通过金融知识体系及专家的经验建立图谱框架，并在框架内填入普适性的实体、关系及属性。后期，随着知识不断更新，人的认知能力不断提升， 图谱将采用自底向上的方式进行更新迭代。
根据本公开提供的实施例，金融数据库可包括用户画像，该方法还可包括：根据用户的属性信息建立或完善用户画像；其中，用户的属性信息包括如下至少之一：年龄因素、家庭收入、可投资金额、家庭负担、投资经验、可接受亏损、心理因素、用户目标数据、用户交互数据、用户行为数据、兴趣爱好。
用户画像的建立可有助于对用户语句进行意图理解，还可以根据用户的风险偏好给出更好的投资建议。
如图16所示，其为本公开实施例提供的确定用户的意图的流程图。根据本公开提供的实施例，根据用户输入的信息，确定用户的意图，可包括步骤501和步骤502。
在步骤501中，根据用户输入的信息，获取关键词和句式。
在步骤502中，基于意图识别规则和分类算法，结合用户画像，按照关键词和句式确定实体和意图。
根据本公开提供的实施例，基于意图识别规则和分类算法，结合用户画像，按照关键词和句式确定实体和意图之前，该方法还可包括：基于关键词和句式确定用户不在执行与投资无关的输入(不在闲聊)，且确定用户不在业务办理流程中。
其中，基于关键词和句式确定用户在执行与投资无关的输入(在闲聊)时，生成闲聊语句或复述语句，将回复反馈至用户。
其中，如果基于关键词和句式确定用户在执行与投资无关的输入(在闲聊)时，若确定用户在询问日期、天气等通用问题，则可以采用通用模板，将回复反馈至用户。
根据本公开提供的实施例，用户输入的信息包括问题时，该方法还可包括：对于多次出现且无法回答的问题，采用阅读理解的方式从金融数据库的数据信息中获取相应的回答。
其中，可基于R-net、SLQA这类模型，通过其实体和意图找出相关文档，并采用阅读理解的方式获取相应回答。
根据本公开提供的实施例，采用阅读理解的方式从金融数据库的数据信息中获取相应的回答之后，该方法还可包括：将审核通过的问题和相应的回答加入金融数据库中的知识库中。
知识库中的QA(问题-回答)对如果涉及金融领域常见实体或事件，可加入金融数据库的图谱库中的知识图谱或事理图谱。
根据本公开提供的实施例，基于金融数据库，按照用户的意图为用户提供相应的金融服务，可包括：按照用户的意图确定用户需要提供金融问题服务时，查询金融数据库中的图谱库，确定存在相应的答案时，输出答案。
用户的意图可以包括需要提供金融问题服务、需要提供数据查询服务、需要业务办理服务、需要投资建议服务等。
根据本公开提供的实施例，查询金融数据库中的图谱库，确定存在相应的答案时，输出答案，可包括：查询图谱库中的知识图谱，确定有相匹配的实体和意图时，通过与金融数据库中的知识库中的映射关系确定相应的答案，输出答案；或者，查询图谱库中的知识图谱，确定没有相匹配的实体和意图时，查询图谱库中的事理图谱，确定有相应的事件，按照事理图谱进行推断分析，输出答案。
根据本公开提供的实施例，基于金融数据库，按照用户的意图为用户提供相应的金融服务，可包括：按照用户的意图确定用户需要提供金融问题服务，基于金融数据库中的图谱库确定没有相应的答案时，基于金融数据库中的知识库，将用户输入的信息与标准问题进行相似度匹配，在相似度大于或等于阈值时，输出标准问题对应的答案。
其中，分别通过传统相似度特征的组合训练冷启动相似度算法，以及根据领域数据训练相应的深度学习模型进行相似度匹配算法；前者具备更强的领域迁移性和鲁棒稳定性，后者在特定领域内拥有更高的精度。
根据本公开提供的实施例，将用户输入的信息与标准问题进行相似度匹配，在相似度小于阈值时，该方法还可包括：确定用户输入的信息中包含图谱库中的实体或意图时，基于实体或意图进行反问；或者，确定用户的问题中不包含图谱库中的实体或意图时，按照预设规则进行通用反问、回复或推荐信息。
根据本公开提供的实施例，基于金融数据库，按照用户的意图为用户提供相应的金融服务，可包括：按照用户的意图确定用户需要提供数据查询服务时，按照用户输入的信息中的关键词查询金融数据库，输出查询结果。
根据本公开提供的实施例，查询结果为数据库没有包含查询内容时，该方法还可包括：确定用户输入的信息中包含金融数据库的图谱库中的实体或意图时，基于实体或意图进行反问；或者，确定用户的问题中不包含图谱库中的实体或意图时，按照预设规则进行通用反问、回复或推荐信息。
根据本公开提供的实施例，基于金融数据库，按照用户的意图为用户提供相应的金融服务，可包括：按照用户的意图确定用户需要业务办理服务时，确定用户需要办理的业务类型，根据业务类型在确定支持办理该业务时，执行业务办理服务。
根据本公开提供的实施例，基于金融数据库，按照用户的意图为用户提供相应的金融服务，可包括：按照用户的意图确定用户需要投资建议服务时，确定用户需要的投资类型；以及，采用元学习的方式，基于金融数据库中的知识库、产品画像和用户画像，为用户进行产品推荐。
如图17所示，其为本公开实施例提供的采用元学习的方式，基于金融数据库中的知识库、产品画像和用户画像，为用户进行产品推荐的流程图。根据本公开提供的实施例，采用元学习的方式，基于金融数据库中的知识库、产品画像和用户画像，为用户进行产品推荐，可包括步骤601-步骤605。
在步骤601中，基于金融数据库中的知识库、产品画像和用户画像，构建环境、行为和状态空间。
在步骤602中，构建优化目标。
在步骤603中，进行任务-行为编码，计算不同状态下不同行为所带来的回报。
在步骤604中，将任务-行为编码嵌入到元-价值网络中，学习任务的损失函数。
在步骤605中，进行策略-梯度训练，优化特定环境及特定状态下的行为，根据优化后的行为，为用户进行产品推荐。
在图17中，通过当前形势分析及事理图谱建立状态空间，再基于每个状态建立行为空间及反馈，最终得到人的最佳行为。
如图18所示，其为本公开实施例提供的深层语义理解及结果反馈的流程图，深层语义理解及结果反馈的流程可包括步骤701-步骤728。
在步骤701中，判断用户是否在闲聊；如果是，用户输入的信息与金融无关，便转入闲聊部分进行处理，执行步骤702；如果否，执行步骤705。
在步骤702中，判断用户是否询问日期、天气等通用问题；如果是，执行步骤703；如果否，执行步骤704。
在步骤703中，采用通用模板，将回复反馈至用户。
在步骤704中，通过语句生成的方式，生成闲聊语句或复述 语句，将回复反馈至用户。
在步骤705中，若不在闲聊，优先判断是否还在流程中；若当前处于某个业务办理流程，执行步骤706；若确认此时不处于业务办理流程，执行步骤708。
在步骤706中，判断是否要终止流程；若是，执行步骤708；若否，执行步骤707。
在步骤707中，引导用户完成该业务办理。
在步骤706～707中，若当前处于某个业务办理流程，则优先引导用户完成该业务办理，除非用户想主动终止该流程。
在步骤708中，基于用户画像和前后文对用户意图进行识别。
在步骤709中，判断意图是否明确，若不明确，执行步骤710；若明确，针对不同类型问题采用不同处理流程；在判断金融问题时，执行步骤711；在判断是数据查询时，执行步骤719；在判断进行业务办理时，执行步骤722；在判断进行投资建议时，执行步骤726。
在步骤710中，进行通用反问，力求做到准确无误。
在步骤711中，查找知识图谱，判断是否有相应的实体和意图；若有，通过与知识库的映射关系找到相应答案，执行步骤718；若没有，执行步骤712。
在步骤712中，查找事理图谱，若有相应事件，根据事理图谱进行推断分析，执行步骤718；若没有，执行步骤713。
在步骤713中，基于知识库与标准问题进行相似度匹配。
在步骤714中，判断是否有大于阈值的标准问题，若是，执行步骤718；若否，执行步骤715。
在步骤715中，从图谱中寻找是否含有相应实体或意图，若有，执行步骤716；若没有，执行步骤717。
在步骤716中，基于图谱中的实体或意图进行反问。
在步骤717中，进行通用反问，告诉用户该问题暂时没有答案，同时可以给用户推荐一些相似度高的热点问题，并继续进行交互。
在步骤718中，给用户相应回复。
在步骤719中，通过关键词提取确定查询内容。
在步骤720中，判断金融数据库中释放包含查询内容，若是执行步骤721；若否，执行步骤715。
在步骤721中，给用户相应结果。用户可通过点击界面获取相关内容。
在步骤722中，判断用户需要办理的业务类型。
在步骤723中，判断是否支持该业务办理，若是，执行步骤724；若否，执行步骤725。
在步骤724中，进入业务办理流程。
在步骤725中，给用户反馈无法办理的回复。
在步骤726中，判断用户需要的投资类型。
在步骤727中，结合元学习模块和产品画像判断最近形式对哪些产品有利。
在步骤728中，结合用户画像给用户投资建议。
本流程中不涉及长文本算法的调用，原因是长文本处理时间过长，会导致用户等待，从而影响体验。长文本处理工作在离线完成，分析结果存储在金融数据库中，便于上述流程调用。
下面以一些具体实例对本公开实施例提供的智能投顾的实现方法进行说明。
实例1：银行智能客服系统
该系统可用于银行虚拟客服中。与传统智能客服系统不同的是，该系统可更好的回答金融相关问题，或是给出投资建议。对于不同用户，可结合用户多维度的画像给出更好的意图理解。针对银行业务领域时，知识库及图谱中要加入相关内容。如图19所示，其为本公开实施例提供的智能投顾的实现方法的另一种流程图。该方法可包括步骤801-步骤806。
在步骤801中，根据用户的个人信息、历史存贷款及购买理财产品的行为，结合预设的问题，初步建立用户画像。
在步骤802中，判断用户是否为闲聊，若是，则进入闲聊模块；若不是，则判断是否正在进行业务流程。如用户说：“我想办理信用卡”，则判断其不属于闲聊。
在步骤803中，如果在进行业务流程，则引导用户完成业务办理；若没有，则通过用户画像理解用户意图。对于之前的问题，会自动转入意图识别。
在步骤804中，基于用户画像和上下文进行意图识别，若意图明确，则继续流程；如果不明确，则进行反问。对于“我想办理信用卡”，意图明确，系统识别为业务办理，则开始业务办理流程。
在步骤805中，根据语义模板，相似度计算等方法识别用户要办理的业务类型。这里，识别为“信用卡办理”业务，银行支持该业务办理。
在步骤806中，按照流程与用户进行交互，引导用户进行业务办理。交互记录将全部保存，方便进行后文语义理解及完善用户画像。
实例2：提供产品资料查询服务
该系统可用于提供产品资料查询服务。如图20所示，其为本 公开实施例提供的智能投顾的实现方法的又一种流程图，该方法可包括步骤901-步骤905。
在步骤901中，根据预设问题建立用户画像，进行意图识别，前几步流程与应用实例1类似。如用户问：“我想查看中兴通讯的股票及其相关研报”，该问题识别为产品资料查询问题，故而走数据查询流程。
在步骤902中，挖掘用户问句中的关键词，采用意图分类方法确定其实体和意图。上述问题中，可以确定实体为“股票”和“研报”，意图为“查询”，限定范围为“中兴通讯”。
在步骤903中，系统接收到消息，若查不到，则基于实体或意图进行反问；若能查到，返回一个链接或是按钮，点击后可进入图形化界面。电脑端是类似wind(金融数据和分析工具服务商)的形式，手机端类似同花顺的股票显示界面。整个界面尽量简洁，仅提供用户想要的信息，如该问题中仅提供股票走势及相关研报列表。当然，用户可以通过界面内的进一步点击或搜索获取其它方面信息。
在步骤904中，若用户关闭界面，则视为用户终止该流程。用户的行为记录将被记录，可用于完善用户画像、便于后文语义理解。
在步骤905中，对于问题中多次出现且系统中无法查询的信息，将会在后台记录，经人工审核后可加入新的数据源。
实例3：提供金融市场分析服务
该系统可用于提供金融市场分析服务。如图21所示，其为本公开实施例提供的智能投顾的实现方法的再一种流程图，该方法可包括步骤1001-步骤1007。
在步骤1001中，根据预设问题建立用户画像，进行意图识别。 前几步流程与应用实例1类似。如用户问：“海啸席卷上海时，会产生什么影响”，该问题识别为金融相关问题，故而走金融问题流程。
在步骤1002中，挖掘用户问句中的关键词，采用意图分类方法确定其实体和意图。如该问题中，实体为“海啸”、“上海”，意图为“影响”。
在步骤1003中，查找知识图谱，是否能找到对应的实体意图组合，若能则返回答案；若不能则进入下一环节。上述问题由于不属于传统金融知识点，且对应不到特定行业或是产品，故而知识图谱中无法得到答案。
在步骤1004中，查找事理图谱，确认是否能找到对应的事件。上述问题属于“上海发生自然灾害”这个事件，若已存入事理图谱，则可以返回该事件发生后会带来的一系列后果。这里假设该事件没有存入事理图谱，则进入相似度计算环节。
在步骤1005中，与知识库中的标准问题进行相似度计算。若匹配上标准问题，则基于QA对返回相应结果；若没有，则进入反问环节。相似度计算主要分为两类，传统特征组合的方式可用于冷启动，当数据逐渐增加，可采用深度学习训练改进的DSSM(Deep Structured Semantic Model，深层结构语义模型)提升相似度匹配效果。
在步骤1006中，判断用户问题是否包含知识图谱中存在的实体或是意图；若存在，则基于该实体或是意图进行反问；若不存在，则进行通用反问。如上述问题中，若知识图谱中存在“上海”这个实体，则反问：“您是想问关于上海的xxx问题吗？”所反问问题尽量与最近发生的事件相关，事件基于图3进行提取，并通过图9建立其与各个行业的联系。
在步骤1007中，对于多次出现且系统暂时无法回答的问题， 将会在后台记录，并会离线通过图5的方式给出相应回答。这类回复中质量较高的将通过人工审核判断其加入图谱还是知识库中。
实例4：提供投资建议
该系统可用于提供投资建议。如图22所示，其为本公开实施例提供的智能投顾的实现方法的再一种流程图，该方法可包括步骤1101-步骤1105。
在步骤1101中，根据预设问题建立用户画像，进行意图识别。前几步流程与应用实例1类似。如用户问：“我想购买理财产品，买什么合适？”，该问题识别为需要提供投资建议，故而走投资建议流程。
在步骤1102中，挖掘用户问句中的关键词，采用意图分类方法确定其实体和意图。如上述问题中，实体为“理财产品”，意图为“购买”，则判断用户想购买理财产品。
在步骤1103中，基于图9的模型，判断近期的各类文本新闻及宏观数据对哪些行业有利(这部分结果会在离线计算并存储)，并基于产品库找出相关的理财产品，并评估其相关程度。
在步骤1104中，基于用户画像给出相应的产品推荐，如用户厌恶风险，则主要推荐低风险、低收益的产品。
在步骤1105中，用户可基于推荐的产品进行自主组合，用户行为将会在后台记录。用户也可以对推荐结果进行反馈(星级评定)，从而进一步完善用户画像。
实例5：提供数据服务
对于资深金融行业从业者，他们更需要基于完备、即时的数据源进行自主分析判断，而不是看系统的分析结果。该系统可直接提供底层结构化数据及图谱查询服务。如图23所示，其为本公开实施例提供的智能投顾的实现方法的再一种流程图，该方法可 包括步骤1201-步骤1205。
在步骤1201中，界面将提供数据库、知识图谱及事理图谱查询的选项，用户可通过点击进入。也可以通过外部接口进行调用。
在步骤1202中，数据库中包含结构化文本数据和宏观市场数据。前者为根据图3处理后的结果，后者包括各类金融产品走势、宏观指数等。
在步骤1203中，知识图谱中存储各类金融知识点，除了宏观金融知识，各类行业之间的联系也会记录在图谱中。
在步骤1204中，事理图谱将记录典型的金融行为过程，主要包含一些周期性金融事件。该部分会随着系统的自学习不断完善。
在步骤1205中，用户对数据库的调用过程同样会在后台记录，进一步完善用户画像。
应用实例6：文本结构化处理
这里着重介绍多元异构数据源获得的数据如何进行结构化处理并存入数据库的流程。如图24所示，其本公开实施例提供的对来自多元异构数据源的数据进行处理的另一种流程图，该流程可包括步骤1301-步骤1309。
在步骤1301中，首先对从数据源获取的文件进行解析，如html文件采用html解析器，pdf文件采用pdf解析器，通过解析器提取文件中有用的文本及图片。
在步骤1302中，进入文本分析器，首先进行文本清洗，然后通过章节分析方法进行粗粒度文本分类。
在步骤1303中，对文本进行句法分析，提取相关的段落及句法特征。
在步骤1304中，对文本中的表情进行识别，转换为文字，同时基于此进行文本情绪分析。
在步骤1305中，提取文件中的图片，并采用图片语义理解的方法对其进行分析(RCNN+RNN)。
在步骤1306中，进入信息抽取器，首先采用规则及深度学习方法对文本进行命名实体识别，提取相应的实体。
在步骤1307中，对文本进行关系及事件抽取，主要采用动态卷积网络，强化学习等方法。
在步骤1308中，对解析得到的表格文件进行结构分析，然后通过表格对齐和补全，抽取完整的表格信息。
在步骤1309中，将清洗后的文本数据和上述抽取的特征及信息，分类存入结构化数据库中。
实例7：行业分析
如图25所示，其为本公开实施例提供的智能投顾的实现方法的再一种流程图，该方法可包括步骤1401-步骤1405。
在步骤1401中，从数据库分别获取结构化文本数据、宏观市场数据及图片数据，根据要处理的问题类别选取不同时间跨度的数据集。如：“近期钢铁行业需求将如何变化？”，则只需读取近期的数据即可。
在步骤1402中，进行多模态数据的特征提取，针对不同类型的数据采用不同特征提取方法。对于长文本数据，除了传统特征和关键词特征，还需要对文本进行摘要，然后将摘要后文本的句向量进行特征提取。
在步骤1403中，将多模态数据的特征嵌入，这里有多种方式，如拼接、归一化后加权和、注意等等。再将嵌入后的特征进行全连接。
在步骤1404中，将上一层的结果处理后与每个行业对应的强关联特征进行嵌入，这里的强关联特征可以采用图3的方法针对 特定行业抽取得到，种类不限。如：钢铁行业上市公司的财报数据、钢铁板块研报中提取的事件、股吧中钢铁板块人们的情绪等等。不同类型的数据有其对应的特征提取方式，具体与1402类似。
在步骤1405中，将1404得到的结果进行处理，激活后导入损失函数层，优化目标为所有损失函数的组合。对于该联合优化问题，每个行业指标数量不同，类型不同，统一根据时间跨度进行归类，如长期、中期、短期，根据不同跨度的指标训练不同的联合学习模型。
实例8：推荐优化
如图26所示，其为本公开实施例提供的智能投顾的实现方法的再一种流程图，该方法可包括步骤1501-步骤1505。
在步骤1501中，构建环境、行为和状态空间。环境和行为如当前形势对某些行业的某些指标有利时，给用户推荐的各类产品组合，状态指用户已持有哪些产品、有多少流动资产等。
在步骤1502中，构建优化目标。如用户年化期望收益。当然，当系统用户量巨大，给用户推荐会影响整个市场时，在保证每个用户达到某个收益的前提下，以全局优化为目标更为合理。
在步骤1503中，进行任务-行为编码，计算不同状态下不同行为所带来的回报。
在步骤1504中，将任务-行为编码嵌入到元-价值网络中，学习整个任务的损失函数。
在步骤1505中，进行策略-梯度训练，优化特定环境及状态下的行为。
如图27所示，本公开实施例还提供一种智能投顾系统，该系统可包括存储器1601、处理器1602及存储在存储器1601上并可在处理器1602上运行的计算机程序1603，处理器1602执行程序 时实现本公开实施提供的智能投顾的实现方法。
本公开实施例还提供一种计算机可读存储介质，存储有计算机可执行指令，计算机可执行指令用于执行本公开实施提供的智能投顾的实现方法。
在本实施例中，上述存储介质可以包括但不限于：U盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。
本领域普通技术人员可以理解，上文中所公开方法中的全部或某些步骤、系统、装置中的功能模块/单元可以被实施为软件、固件、硬件及其适当的组合。在硬件实施方式中，在以上描述中提及的功能模块/单元之间的划分不一定对应于物理组件的划分；例如，一个物理组件可以具有多个功能，或者一个功能或步骤可以由若干物理组件合作执行。某些组件或所有组件可以被实施为由处理器，如数字信号处理器或微处理器执行的软件，或者被实施为硬件，或者被实施为集成电路，如专用集成电路。这样的软件可以分布在计算机可读介质上，计算机可读介质可以包括计算机存储介质(或非暂时性介质)和通信介质(或暂时性介质)。如本领域普通技术人员公知的，术语计算机存储介质包括在用于存储信息(诸如计算机可读指令、数据结构、程序模块或其他数据)的任何方法或技术中实施的易失性和非易失性、可移除和不可移除介质。计算机存储介质包括但不限于RAM、ROM、EEPROM、闪存或其他存储器技术、CD-ROM、数字多功能盘(DVD)或其他光盘存储、磁盒、磁带、磁盘存储或其他磁存储装置、或者可以用于存储期望的信息并且可以被计算机访问的任何其他的介质。此外，本领域普通技术人员公知的是，通信介质通常包含计算机可读指令、数据结构、程序模块或者诸如载波或其他传输 机制之类的调制数据信号中的其他数据，并且可包括任何信息递送介质。

Claims (25)

  1. 一种智能投顾的实现方法，包括：
    基于自然语言处理NLP对来自多元异构数据源的数据进行处理，并将得到的数据信息加入金融数据库；
    根据用户输入的信息，确定所述用户的意图；以及
    基于所述金融数据库，根据所述用户的意图为所述用户提供金融服务。
  2. 如权利要求1所述的方法，其中，基于NLP对所述来自多元异构数据源的数据进行处理，包括：
    对所述来自多元异构数据源的数据进行数据解析，得到文本数据；
    对所述文本数据进行文本分析，得到文本信息；以及
    根据所述文本信息进行信息抽取，得到结构化文本数据。
  3. 如权利要求2所述的方法，在根据所述文本信息进行信息抽取之前，还包括：对所述文本信息进行摘要处理。
  4. 如权利要求1所述的方法，其中，所述数据信息包括多模态数据，所述方法还包括：
    从所述金融数据库中获取所述多模态数据；
    基于多模态输入的联合学习模型，建立所述多模态数据与行业之间的关联关系；以及
    将所述多模态数据与行业之间的关联关系存入所述金融数据库的知识库中。
  5. 如权利要求4所述的方法，其中，所述多模态数据包括结构化文本数据、宏观市场数据和图片数据，基于所述多模态输入的联合学习模型，建立所述多模态数据与行业之间的关联关系， 包括：
    对所述结构化文本数据、所述宏观市场数据以及所述图片数据进行特征提取；
    将所述结构化文本数据的特征、所述宏观市场数据的特征以及所述图片数据的特征进行嵌入，并将嵌入后的特征进行全连接；以及
    将全连接的结果与每个行业对应的强关联特征进行嵌入，进行联合优化，建立所述多模态数据与行业之间的关联关系。
  6. 如权利要求4所述的方法，还包括：
    基于元学习的方式建立注意力模型；以及
    根据所述注意力模型优化所述联合学习模型。
  7. 如权利要求1所述的方法，其中，所述金融数据库包括图谱库和知识库，所述图谱库包含一个知识图谱和多个事理图谱，且所述知识图谱与所述知识库相关联。
  8. 如权利要求7所述的方法，还包括：将所述知识库中涉及金融领域的实体或事件加入所述金融数据库的图谱库中。
  9. 如权利要求1所述的方法，其中，根据所述用户输入的信息，确定所述用户的意图，包括：
    根据所述用户输入的信息，获取关键词和句式；以及
    基于意图识别规则和分类算法，根据用户画像、所述关键词和所述句式，确定实体和所述用户的意图。
  10. 如权利要求1所述的方法，其中，所述用户输入的信息包括问题，所述方法还包括：对于多次出现且无法回答的问题，采用阅读理解的方式从所述金融数据库的所述数据信息中获取相应的第一答案。
  11. 如权利要求10所述的方法，在采用所述阅读理解的方式从所述金融数据库的所述数据信息中获取所述相应的第一答案之后，还包括：将审核通过的所述问题和所述相应的第一答案加入所述金融数据库的知识库中。
  12. 如权利要求1所述的方法，其中，基于所述金融数据库，根据所述用户的意图为所述用户提供金融服务，包括：
    响应于根据所述用户的意图确定所述用户需要提供金融问题服务，查询所述金融数据库的图谱库，确定存在相应的第二答案；以及
    输出所述第二答案。
  13. 如权利要求12所述的方法，其中，查询所述金融数据库的图谱库，确定存在所述相应的第二答案，包括：
    响应于查询所述图谱库的知识图谱，确定有相匹配的实体和意图，通过与所述金融数据库的知识库中的映射关系确定所述相应的第二答案；或者
    响应于查询所述图谱库的知识图谱，确定没有相匹配的实体和意图，查询所述图谱库的事理图谱，确定相应的事件，根据所述事理图谱进行推断分析，确定所述相应的第二答案。
  14. 如权利要求1所述的方法，其中，基于所述金融数据库，根据所述用户的意图为所述用户提供金融服务，包括：
    响应于根据所述用户的意图确定所述用户需要提供金融问题服务，且基于所述金融数据库的图谱库确定没有相应的答案，基于所述金融数据库的知识库，将所述用户输入的信息与标准问题进行相似度匹配；以及
    响应于确定所述用户输入的信息与设定标准问题的相似度大于或等于阈值，输出与所述设定标准问题对应的第三答案。
  15. 如权利要求14所述的方法，还包括：
    响应于确定所述用户输入的信息与所有标准问题的相似度小于所述阈值，且确定所述用户输入的信息中包含所述图谱库中的实体或意图，基于所述用户输入的信息中包含的所述图谱库中的实体或意图进行反问；或者
    响应于确定所述用户输入的信息与所有标准问题的相似度小于所述阈值，且确定所述用户输入的信息中不包含所述图谱库中的实体或意图，根据第一预设规则进行通用反问、回复或推荐信息。
  16. 如权利要求1所述的方法，其中，基于所述金融数据库，根据所述用户的意图为所述用户提供金融服务，包括：
    响应于根据所述用户的意图确定所述用户需要提供数据查询服务，根据所述用户输入的信息中的关键词查询所述金融数据库，并输出查询结果。
  17. 如权利要求16所述的方法，还包括：
    响应于所述查询结果为所述金融数据库中没有包含待查询内容，且确定所述用户输入的信息中包含所述金融数据库的图谱库中的实体或意图，基于所述用户输入的信息中包含的所述金融数据库的图谱库中的实体或意图进行反问；或者
    响应于所述查询结果为所述金融数据库中没有包含待查询内容，且确定所述用户输入的信息中不包含所述金融数据库的图谱库中的实体或意图，根据第二预设规则进行通用反问、回复或推荐信息。
  18. 如权利要求1所述的方法，其中，基于所述金融数据库，根据所述用户的意图为所述用户提供金融服务，包括：
    响应于根据所述用户的意图确定所述用户需要业务办理服务， 确定所述用户需要办理的业务类型；
    根据所述业务类型确定支持办理所述业务；以及
    执行所述业务的办理。
  19. 如权利要求1所述的方法，其中，基于所述金融数据库，根据所述用户的意图为所述用户提供金融服务，包括：
    响应于根据所述用户的意图确定所述用户需要投资建议服务，确定所述用户需要的投资类型；以及
    采用元学习的方式，基于所述金融数据库的知识库、产品画像和用户画像，为所述用户推荐产品。
  20. 如权利要求19所述的方法，其中，采用元学习的方式，基于所述金融数据库的所述知识库、所述产品画像和所述用户画像，为所述用户推荐产品，包括：
    基于所述金融数据库的所述知识库、所述产品画像和所述用户画像，构建环境、行为和状态空间；
    构建优化目标；
    进行任务-行为编码，计算不同状态下不同行为所带来的回报；
    将所述任务-行为编码嵌入到元-价值网络中，学习任务的损失函数；以及
    进行策略-梯度训练，优化特定环境及特定状态下的行为，根据优化后的行为，为所述用户推荐产品。
  21. 一种智能投顾系统，包括：数据处理模块、深层语意理解模块、结果反馈模块和金融数据库，其中：
    所述数据处理模块，配置为基于自然语言处理NLP对来自多元异构数据源的数据进行处理，并将得到的数据信息加入所述金融数据库；以及，获取并处理用户输入的信息，将处理后的信息 发送至所述深层语意理解模块；
    所述深层语意理解模块，配置为根据所述处理后的信息，确定所述用户的意图；以及
    所述结果反馈模块，配置为基于所述金融数据库，根据所述用户的意图为所述用户提供金融服务。
  22. 如权利要求21所述的智能投顾系统，其中，所述数据信息包括多模态数据，所述智能投顾系统还包括：
    联合学习模块，配置为从所述金融数据库中获取所述多模态数据；基于多模态输入的联合学习模型，建立所述多模态数据与行业之间的关联关系；以及，将所述多模态数据与行业之间的关联关系存入所述金融数据库中的知识库中。
  23. 如权利要求22所述的智能投顾系统，还包括：
    元学习模块，配置为基于元学习的方式建立注意力模型；以及，根据所述注意力模型优化所述联合学习模型。
  24. 一种智能投顾系统，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，其中，所述处理器执行所述程序时实现如权利要求1～20中任一项所述的智能投顾的实现方法。
  25. 一种计算机可读存储介质，其上存储有一个或者多个计算机程序，所述一个或者多个计算机程序可被一个或者多个处理器执行，以实现如权利要求1～20中任一项所述的智能投顾的实现方法。
PCT/CN2019/120573 2018-12-05 2019-11-25 一种智能投顾的实现方法及系统 WO2020114269A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811481844.7 2018-12-05
CN201811481844.7A CN111353013A (zh) 2018-12-05 2018-12-05 一种智能投顾的实现方法及系统

Publications (1)

Publication Number Publication Date
WO2020114269A1 true WO2020114269A1 (zh) 2020-06-11

Family

ID=70975111

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/120573 WO2020114269A1 (zh) 2018-12-05 2019-11-25 一种智能投顾的实现方法及系统

Country Status (2)

Country Link
CN (1) CN111353013A (zh)
WO (1) WO2020114269A1 (zh)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112241631A (zh) * 2020-10-23 2021-01-19 平安科技（深圳）有限公司 文本语义识别方法、装置、电子设备及存储介质
CN112463841B (zh) * 2020-11-03 2023-04-21 贵州江南航天信息网络通信有限公司 一种基于工业大数据的智能化决策与精准推送的方法及引擎
CN112347314B (zh) * 2020-11-20 2023-04-14 国网信通亿力科技有限责任公司 一种基于图数据库的数据资源管理系统
CN112487293A (zh) * 2020-11-30 2021-03-12 增长引擎（北京）信息技术有限公司 一种安全事故案例结构化信息抽取方法、装置及介质
CN112989038B (zh) * 2021-02-08 2022-06-21 浙江连信科技有限公司 句子级用户画像生成方法、装置及存储介质
CN113239679A (zh) * 2021-04-17 2021-08-10 上海快确信息科技有限公司 一种基于深度学习的债券标位自动检测技术
CN113393084B (zh) * 2021-05-13 2024-06-11 上海湃道智能科技有限公司 作业票流程管理系统
CN113641835B (zh) * 2021-08-27 2022-10-11 北京达佳互联信息技术有限公司 多媒体资源推荐方法、装置、电子设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8930247B1 (en) * 2007-01-31 2015-01-06 Financialsharp, Inc. System and methods for content-based financial decision making support
CN105094315A (zh) * 2015-06-25 2015-11-25 百度在线网络技术（北京）有限公司 基于人工智能的人机智能聊天的方法和装置
CN107562856A (zh) * 2017-08-28 2018-01-09 深圳追科技有限公司 一种自助式客户服务系统及方法
CN107958091A (zh) * 2017-12-28 2018-04-24 北京贝塔智投科技有限公司 一种基于金融垂直知识图谱的nlp人工智能方法及交互系统
CN108491386A (zh) * 2018-03-19 2018-09-04 上海携程国际旅行社有限公司 自然语言理解方法及系统

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108121824A (zh) * 2018-01-12 2018-06-05 北京融快线科技有限公司 一种面向金融服务的聊天机器人及系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8930247B1 (en) * 2007-01-31 2015-01-06 Financialsharp, Inc. System and methods for content-based financial decision making support
CN105094315A (zh) * 2015-06-25 2015-11-25 百度在线网络技术（北京）有限公司 基于人工智能的人机智能聊天的方法和装置
CN107562856A (zh) * 2017-08-28 2018-01-09 深圳追科技有限公司 一种自助式客户服务系统及方法
CN107958091A (zh) * 2017-12-28 2018-04-24 北京贝塔智投科技有限公司 一种基于金融垂直知识图谱的nlp人工智能方法及交互系统
CN108491386A (zh) * 2018-03-19 2018-09-04 上海携程国际旅行社有限公司 自然语言理解方法及系统

Also Published As

Publication number Publication date
CN111353013A (zh) 2020-06-30

Similar Documents

Publication Publication Date Title
WO2020114269A1 (zh) 一种智能投顾的实现方法及系统
US11748555B2 (en) Systems and methods for machine content generation
Algaba et al. Econometrics meets sentiment: An overview of methodology and applications
US20210272040A1 (en) Systems and methods for language and speech processing with artificial intelligence
US11347803B2 (en) Systems and methods for adaptive question answering
US20230252224A1 (en) Systems and methods for machine content generation
Chang et al. Accelerating mixed methods research with natural language processing of big text data
US10733619B1 (en) Semantic processing of customer communications
US20230214949A1 (en) Generating issue graphs for analyzing policymaker and organizational interconnectedness
EP3855320A1 (en) Systems and methods for adaptive question answering related applications
US20160283876A1 (en) System and method for providing automomous contextual information life cycle management
US20230214753A1 (en) Generating issue graphs for analyzing organizational influence
Manzoor et al. Towards retrieval-based conversational recommendation
Hartmann et al. Natural language processing in marketing
Anastasiei et al. Automated vs manual content analysis–A retrospective look
Zhang et al. Modeling and prediction of stock price with convolutional neural network based on blockchain interactive information
Bae et al. Mobilizing text as data
Lamba et al. Sentiment analysis
Lima et al. Towards ubiquitous requirements engineering through recommendations based on context histories
Iovine et al. Virtual customer assistants in finance: From state of the art and practices to design guidelines
Pentland et al. Does accuracy matter? Methodological considerations when using automated speech-to-text for social science research
Mazza et al. A text mining approach for CSR communication: an explorative analysis of energy firms on Twitter in the post-pandemic era
US20190294965A1 (en) Systems and methods using artificial intelligence to analyze natural language sources based on personally-developed intelligent agent models
Catanese et al. Natural language processing in official statistics: The social mood on economy index experience
Lo et al. From ELIZA to ChatGPT: The Evolution of NLP and Financial Applications

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19892288

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 15/10/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19892288

Country of ref document: EP

Kind code of ref document: A1