A rapid construction method and lifecycle management platform for financial question-answering algorithms

Abstract

The present invention discloses a method for rapidly constructing a financial question-and-answer algorithm and a lifecycle management platform, relating to the field of artificial intelligence technology, and solving the problems of low efficiency in constructing a financial question-and-answer system and imperfect system lifecycle management. The present invention comprises S1: collecting financial data and performing preprocessing; S2: constructing an artificial intelligence conversation interaction algorithm; and S3: integrating domain expertise into the artificial intelligence conversation interaction algorithm. The present invention realizes automated data collection from multiple sources and angles, and enhances the question-and-answer system's ability to understand and express the correlation between questions and answers. In addition, the present invention realizes rapid construction and lightweighting of the model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for rapidly constructing a financial question-answering algorithm and a lifecycle management platform. Background Art

[0002] In the financial sector, NLP technology is used to process large amounts of financial text data, such as financial news, company reports, and user questions. By combining NLP technology with financial data, financial question-answering systems can be built to help users quickly obtain relevant financial information, conduct data analysis, and provide decision support.

[0003] However, these technologies and methods still have some shortcomings and challenges. Some of these issues include:

[0004] 1. Inefficient algorithm construction: Current financial question-answering system construction typically requires significant time and resources, including data collection, annotation, model training, and optimization. Existing methods have yet to adequately address the need for rapidly building financial question-answering algorithms.

[0005] 2. Incomplete system lifecycle management: In the practical application of financial question-answering systems, model lifecycle management is a critical and complex task. Existing methods lack an integrated management platform to effectively support model generation, testing, deployment, monitoring, and updates.

[0006] 3. Lack of integration of domain expertise: Question answering systems in the financial sector require the ability to understand financial-related questions and domain knowledge. Existing methods are insufficient in integrating domain expertise, limiting the accuracy and effectiveness of question answering systems in the financial sector. Summary of the Invention

[0007] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a method for quickly constructing a financial question-answering algorithm and a lifecycle management platform, which aims to solve the problems of low efficiency in constructing the financial question-answering system and imperfect system lifecycle management.

[0008] A method for quickly constructing a financial question-answering algorithm includes the following steps:

[0009] S1: Collect financial data and preprocess them;

[0010] S2: Build AI conversational interaction algorithms;

[0011] S3: Integrate domain expertise into AI conversational interaction algorithms.

[0012] Preferably, the S1 comprises the following steps:

[0013] S11: Using automated data collection methods based on open source datasets, web crawler technology, and log collection, we can quickly acquire large-scale question-answering data from multiple financial-related data sources, including financial news websites, reports, documents, and user questions.

[0014] S12: Preprocess the collected large-scale question-answering data using natural language processing and text cleaning, including denoising, text normalization, and semantic parsing;

[0015] S13: Extract features of pre-processed data using LoRa technology.

[0016] Preferably, the S2 comprises the following steps:

[0017] S21: Build a bidirectional information decoupled multi-round dialogue network based on reinforcement learning as a universal dialogue encoder; combine autoregressive networks and autoencoder networks for pre-training to continuously improve intent recognition capabilities;

[0018] S22: By integrating pre-trained autoregressive and autoencoder networks with a universal conversational encoder, we build a general-purpose large-scale natural language processing model, while also incorporating large-scale knowledge graph data.

[0019] S23: Optimizing general-domain large-scale natural language processing models via open-domain multi-turn dialogue policy learning.

[0020] Preferably, the formula for the two-way information decoupling multi-round dialogue network combined with reinforcement learning as a universal dialogue encoder in S21 is as follows:

[0021] H t =BIDirectionalRNN(D h ,H t-1 ,Q t )

[0022]

[0023] Where J(θ) represents the objective function, θ represents the model parameters, Q represents the sequence of dialogue history, including the dialogue content of multiple rounds; A represents the action sequence, which represents the encoding selection of the bidirectional information decoupling network in each round; R t represents the reward sequence, which represents the reinforcement learning reward in each round; BIDirectionalRNN represents a bidirectional recurrent neural network for encoding the dialogue history; D h Represents the parameters of the network, which are used to control the encoding process of the bidirectional information decoupling network.

[0024] Preferably, the autoencoder network in S21 is incrementally pre-trained using multi-task learning; specifically:

[0025] The training objective of the autoencoder network can be expressed as:

[0026]

[0027] Among them, L autoencoder represents the training loss of the autoencoder network, X t represents the original input, represents the input for reconstruction;

[0028] The goal of multi-task learning can be expressed as:

[0029]

[0030] Among them, L MTL represents the total loss of multi-task learning, α i Represents the weight of task i, Loss i That is the loss of task i;

[0031] The autoregressive network is based on the Tranformer-XL structure and supports long text language model modeling.

[0032] Preferably, the universal conversation encoder in the general-domain large-scale natural language processing model in S22 can be expressed as:

[0033] Z=Encoder(Q)

[0034] Among them, Z represents the encoded features of the dialogue history, Encoder represents the universal dialogue encoder, and Q represents the sequence of dialogue history.

[0035] Preferably, the S24 includes:

[0036] Build a dialogue engine based on rule-based dialogue flow, implement preliminary modeling of dialogue management based on supervised learning, continuously enhance dialogue capabilities through self-learning algorithms, proactively plan dialogue content, and improve multi-round coherence and controllability;

[0037] The dialogue engine is modeled as a function that generates actions according to rules:

[0038] A t =RuleEngine(D h ,Q t )

[0039] Among them, A t represents the action of round t, RuleEngine represents the rule-based dialogue engine, and D h Represents the encoded features of the conversation history, Q t Indicates the content of the conversation in round t;

[0040] The preliminary model of supervised learning to achieve dialogue management is expressed as:

[0041] P(A t |D h ,, Q t)=Softmax(W·[D h ;Q t ])

[0042] Among them, P(A t |Q h ,Q t ) represents the predicted action probability distribution, W represents the model parameters, [D h ;Q t ] represents the connection between the encoding features of the conversation history and the current conversation content;

[0043] The self-learning algorithm updates the dialogue management model through a reinforcement learning framework:

[0044]

[0045] The active planning of dialogue content and improvement of multi-round coherence and controllability are expressed as follows:

[0046] Q t+1 =Planner(D h ,Q t ,A t )

[0047] Among them, Q t+1 represents the content of the conversation in round t+1, Planner represents the active planning strategy, and D h Represents the encoded features of the conversation history, Q t Indicates the content of the conversation in round t, A t Represents the action in round t.

[0048] Preferably, the S3 comprises the following steps:

[0049] S31: Constructing a financial knowledge graph;

[0050] S32: Build an application named entity recognition model;

[0051] S33: Applying financial knowledge graphs and entity recognition to the training and reasoning of artificial intelligence conversational interaction algorithm models.

[0052] Preferably, the S31 includes:

[0053] The financial knowledge graph includes the definition of entity nodes and relationship edges, where the entity set E represents various entities in the financial field, including companies, indicators, and terms. The relationship set R represents the relationship between entities, including upstream and downstream relationships and similar relationships. The adjacency matrix A is used to represent the relationship of the graph. Specifically:

[0054]

[0055] Among them A ij Indicates whether there is a relationship between entity i and entity j;

[0056] The S32 includes:

[0057] In the process of named entity recognition, the conditional random field CRF model is used to identify entities; given a text sequence T, a state sequence S is defined, where S i Indicates whether position i is a named entity; the NER module algorithm formula is as follows:

[0058]

[0059] Among them, Z(T) is the normalization factor, n is the length of the text sequence, k is the number of feature functions, and λ j is the feature weight, f j (i, T, S) is the jth feature function, which represents the relationship between the feature at position i and the state S;

[0060] The S33 includes:

[0061] Incorporating professional knowledge into AI conversational interaction algorithm models to improve the system's understanding and accuracy of financial questions. This includes: combining entity recognition results with knowledge graph information to match and reason between questions and answers; using knowledge graphs and entity recognition results to enhance the understanding capabilities of AI conversational interaction algorithm models; calculating the correlation between entities in the question and entities in the knowledge graph, and then matching questions and answers based on this correlation.

[0062] Correlation (e i ,e j )=similarity(e i ,e j )×weight(e j )

[0063] Among them, e i and e j Represent the entities in the question and the entities in the knowledge graph respectively, and the similarity (e i ,e j ) represents the similarity between entities, and the weight (e j) represents the weight of the entity in the knowledge graph; then the entity with the highest correlation is selected for matching and reasoning between questions and answers.

[0064] A lifecycle management platform for financial question-answering algorithms, the platform is used to manage financial question-answering algorithms, including the generation, testing, deployment, monitoring, and updating of algorithm models; the system lifecycle management platform includes:

[0065] Graph databases facilitate users to store model parameters, dataset information, and entity relationships, enabling efficient query and retrieval capabilities;

[0066] A visual operation and configuration interface is used to configure algorithm model hyperparameters, select data sets, and set monitoring indicators; an automated model update and version control mechanism is used to regularly detect changes in data sets, monitor the performance of algorithm models, and automatically store model retraining and deployment according to configuration parameters to ensure that the algorithm model is in the optimal state.

[0067] The beneficial effects of the present invention include:

[0068] a. Automated data collection from multiple sources and angles.

[0069] b. Transfer learning of pre-trained language models and lightweight model design: This invention combines transfer learning of pre-trained language models, RNN or Transformer models with attention mechanisms, and lightweight model design. Through transfer learning of pre-trained language models and model lightweighting technology, the model training speed and reasoning efficiency are accelerated. At the same time, the introduction of the attention mechanism enables the model to automatically focus on important information and perform weighted processing, thereby improving the question-answering system's ability to understand and express the correlation between questions and answers. In addition, knowledge distillation technology is used to transfer the knowledge of complex models to simplified models, enabling rapid model construction and lightweighting.

[0070] c. Sequence-to-sequence model with attention mechanism: By adopting a sequence-to-sequence (Seq2Seq) model and introducing an attention mechanism, the model can automatically focus on important information between questions and answers and perform weighted processing. This integration of algorithmic models improves the question-answering system's ability to understand and express the relationship between questions and answers.

[0071] d. Structured matching in financial question-answering scenarios: using named entity recognition (NER) + category prediction (intent recognition)

[0072] + bert4keras implements k-bert; the present invention constructs a financial knowledge graph and applies entity recognition models, such as named entity recognition (NER), to identify entities related to the financial field; by applying knowledge graphs and entity recognition to the training and reasoning process of question-answering algorithm models, the system can more accurately understand the proper nouns, terms, and entities in the financial field, improving the accuracy and effectiveness of question-answering systems in the financial field.

[0073] e. Integrated system lifecycle management platform: This invention designs and implements an integrated system lifecycle management platform that covers model generation, testing, deployment, monitoring, and updating. By introducing automated model updates and version control mechanisms, as well as a visual user interface, this invention makes the management of financial question-and-answer systems more convenient and efficient. The platform, combined with the application of a graph database, provides efficient model management, knowledge query, and reasoning capabilities. This innovative system lifecycle management platform promotes system stability, reliability, and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A schematic diagram of a method for rapidly constructing a financial question-answering algorithm according to Example 1

[0075] Figure 2 A schematic diagram of the artificial intelligence conversation interaction algorithm involved in Example 1.

[0076] Figure 3 This is the overall structure diagram of named entity recognition + category prediction + bert4keras involved in Example 1.

[0077] Figure 4 This is the overall structure diagram of the lifecycle management platform involved in Example 2. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0079] Example 1

[0080] The following is combined with Figure 1-4 The specific embodiments of the present invention are described in detail;

[0081] A method for quickly constructing a financial question-answering algorithm includes the following steps:

[0082] S1: collect data and perform preprocessing;

[0083] Automated data collection methods based on open-source datasets, web crawler technology, and log collection are employed to rapidly acquire large-scale question-and-answer data from multiple financial data sources. Furthermore, Chinese text cleaning and semantic parsing technology based on HeackMT5-ZhCleanText1ML is introduced. Combined with LoRa technology, this approach maintains the effectiveness of existing model training, prevents context loss, and achieves results comparable to full-scale training. Preprocessing steps remove noise, standardize the data format, and perform semantic analysis to improve data quality and usability, significantly enhancing algorithm development efficiency and data preparation reliability. Furthermore, data augmentation techniques are employed to expand the training dataset, further improving the model's generalization and robustness.

[0084] S2: Build AI conversational interaction algorithms, such as Figure 2 As shown;

[0085] S21: Building a two-way information decoupling multi-round dialogue network based on reinforcement learning;

[0086] Combine the bidirectional information decoupling network with reinforcement learning as a universal dialogue encoder:

[0087] H t =BIDirectionalRNN(D h ,H t-1 ,Q t )

[0088]

[0089] Where J(θ) represents the objective function, θ represents the model parameters, Q represents the sequence of dialogue history, including the dialogue content of multiple rounds; A represents the action sequence, which represents the encoding selection of the bidirectional information decoupling network in each round; R t represents the reward sequence, which represents the reinforcement learning reward in each round; BIDirectionalRNN represents a bidirectional recurrent neural network for encoding the dialogue history; D h Represents the parameters of the network, which are used to control the encoding process of the bidirectional information decoupling network.

[0090] S22: Continuously improve intent recognition capabilities based on a multi-paradigm unified pre-trained large model;

[0091] The autoregressive and autoencoder networks are combined for pre-training. The autoencoder network is incrementally pre-trained using multi-task learning. Specifically:

[0092] The training objective of the autoencoder network can be expressed as:

[0093]

[0094] Among them, L autoencoder represents the training loss of the autoencoder network, X t represents the original input, represents the input for reconstruction;

[0095] The goal of multi-task learning can be expressed as:

[0096]

[0097] Among them, L MTL represents the total loss of multi-task learning, α i Represents the weight of task i, Loss i That is the loss of task i;

[0098] During the training of the autoencoder network, the reconstructed input It can include information from different pre-training tasks. This can be achieved by fusing features from different tasks into the reconstruction process, thereby forcing the model to learn more semantic information during the pre-training stage.

[0099] The objective function L of multi-task learning MTL The loss term Loss in i Represents losses for different tasks. These task losses can be combined during pre-training to establish a shared semantic representation in the model. The reconstruction process of the autoencoder network can increase task relevance by minimizing one of these task losses, thereby fusing multiple semantic information into the model.

[0100] In this way, the connection between autoencoder networks and multi-task learning can achieve more powerful pre-trained models, thereby continuously improving intent recognition capabilities.

[0101] The autoregressive network is based on the Tranformer-XL structure and supports long text language model modeling.

[0102] S23: Build a large-scale natural language processing model for general domains;

[0103] By integrating autoregressive networks and autoencoder networks, a general-purpose large-scale natural language processing model is built, and large-scale knowledge graph data is introduced. The general conversation encoder in the general-purpose large-scale natural language processing model can be expressed as:

[0104] Z=Encoder(Q)

[0105] Among them, Z represents the encoded features of the dialogue history, Encoder represents the universal dialogue encoder, and Q represents the sequence of dialogue history.

[0106] The general-purpose large-scale natural language processing model is a holistic architecture designed to handle a variety of natural language processing tasks. The general conversation encoder is part of this model and is responsible for handling conversation-related tasks such as identifying conversational intent and generating responses. Large-scale knowledge graph data is introduced to enrich the model's common sense knowledge to better handle various tasks. This knowledge is incorporated into the general conversation encoder to improve the model's ability to answer common sense questions from the conversation history.

[0107] S24: Research on open domain multi-round dialogue strategy learning and optimization technology;

[0108] Build a dialogue engine based on rule-based dialogue flow, implement preliminary modeling of dialogue management based on supervised learning, continuously enhance dialogue capabilities through self-learning algorithms, proactively plan dialogue content, and improve multi-round coherence and controllability;

[0109] The dialogue engine is modeled as a function that generates actions according to rules:

[0110] A t =RuleEngine(D h ,Q t )

[0111] Among them, A t represents the action of round t, RuleEngine represents the rule-based dialogue engine, and D h Represents the encoded features of the conversation history, Q t Represents the content of the conversation in round t.

[0112] The preliminary model of supervised learning to achieve dialogue management is expressed as:

[0113] P(A t |D h ,Q t )=Softmax(W·[D h ;Q t ])

[0114] Among them, P(A t |D h ,Q t ) represents the predicted action probability distribution, W represents the model parameters, [D h ;Q t ] represents the connection between the conversation history encoding features and the current conversation content.

[0115] The self-learning algorithm updates the dialogue management model using a reinforcement learning framework, with the parameters described above:

[0116]

[0117] The active planning of dialogue content and improvement of multi-round coherence and controllability are expressed as follows:

[0118] Q t+1 =Planner(D h ,Q t ,A t )

[0119] Among them, Q t+1 represents the content of the conversation in round t+1, Planner represents the active planning strategy, and D h Represents the encoded features of the conversation history, Q t Indicates the content of the conversation in round t, A t Represents the action in round t.

[0120] S3: Integrate domain expertise into AI conversational interaction algorithms, e.g. Figure 3 As shown;

[0121] S31: Constructing a financial knowledge graph;

[0122] The financial knowledge graph includes the definition of entity nodes and relationship edges, where the entity set E represents various entities in the financial field, including companies, indicators, and terms. The relationship set R represents the relationship between entities, including upstream and downstream relationships and similar relationships. The adjacency matrix A is used to represent the relationship of the graph. Specifically:

[0123]

[0124] Among them A ij Indicates whether there is a relationship between entity i and entity j.

[0125] S32: Build an application named entity recognition model;

[0126] In the process of named entity recognition, the conditional random field CRF model is used to identify entities. Given a text sequence T, a state sequence S is defined, where S i Indicates whether the position i is a named entity. The NER module algorithm formula is as follows:

[0127]

[0128] Among them, Z(T) is the normalization factor, n is the length of the text sequence, k is the number of feature functions, and λ j is the feature weight, f j(i, T, S) is the jth feature function, which represents the relationship between the feature at position i and the state S.

[0129] S33: Applying Financial Knowledge Graph and Entity Recognition to Model Training and Inference

[0130] Incorporating professional knowledge into the question-answering algorithm model improves the system's understanding and accuracy of financial questions. For example, during the model inference phase, entity recognition results and knowledge graph information are combined to match and infer questions and answers. During model training and inference, the knowledge graph and entity recognition results are used to enhance the understanding capabilities of the question-answering algorithm model. For example, during the inference phase, the correlation between entities in the question and entities in the knowledge graph can be calculated, and then the question and answer can be matched based on this correlation.

[0131] Correlation (e i ,e j )=Similarity(e i ,e j )×weight(e j )

[0132] Among them, e i and e j Represent the entities in the question and the entities in the knowledge graph respectively, and the similarity (e i ,e j ) represents the similarity between entities, and the weight (e j ) represents the weight of the entity in the knowledge graph. Then the entity with the highest relevance is selected for question-answer matching and reasoning.

[0133] The improvement data of human-computer dialogue performance index obtained by using the method of this embodiment is shown in the following table:

[0134] Table 1 Improvement of human-computer dialogue performance indicators

[0135] index Traditional methods This method Percent improvement Response accuracy 75％ 81％ +6％ Response recall rate 75％ 83％ +8％

[0136] Example 2

[0137] like Figure 4 As shown, a life cycle management platform for financial question answering algorithms should be understood as follows: Figure 1 The method embodiment is used to manage the financial question-answering algorithm constructed in Example 1, and the management includes the generation, testing, deployment, monitoring and updating of the algorithm model; the system lifecycle management platform includes:

[0138] Graph databases facilitate users to store model parameters, dataset information, and entity relationships, enabling efficient query and retrieval capabilities;

[0139] A visual operation and configuration interface for configuring algorithm model hyperparameters, selecting data sets, and setting monitoring indicators;

[0140] The automated model update and version control mechanism is used to regularly detect changes in the data set, monitor the performance of the algorithm model, and automatically store the model for retraining and deployment based on the configuration parameters, thereby ensuring that the algorithm model is in the optimal state.

[0141] The system lifecycle management platform performance data obtained by using the method of this embodiment is shown in the following table:

[0142] Table 2 Effect of system life cycle management platform

[0143] Effect data Reduced failure rate 50％ Reduced maintenance workload 30％

[0144] The above-described embodiments merely represent specific implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of the present application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the technical concept of the present application, and all such variations and improvements fall within the scope of protection of the present application.

Claims

1. A method for rapidly constructing a financial question-answering algorithm, characterized in that: The following steps are involved: S1: Collect financial data and preprocess them; S2: Build AI conversational interaction algorithms; S3: Integrate domain expertise into AI conversational interaction algorithms; Described S2 comprises the following steps: S21: Build a bidirectional information decoupled multi-round dialogue network based on reinforcement learning as a universal dialogue encoder; combine autoregressive networks and autoencoder networks for pre-training to continuously improve intent recognition capabilities; S22: By integrating pre-trained autoregressive and autoencoder networks with a universal conversational encoder, we build a general-purpose large-scale natural language processing model, while also incorporating large-scale knowledge graph data. S23: Optimizing general-domain large-scale natural language processing models through open-domain multi-turn dialogue strategy learning; S24: Research on open domain multi-round dialogue strategy learning and optimization technology; The S24 includes: Build a dialogue engine based on rule-based dialogue flow, implement preliminary modeling of dialogue management based on supervised learning, continuously enhance dialogue capabilities through self-learning algorithms, proactively plan dialogue content, and improve multi-round coherence and controllability; The dialogue engine is modeled as a function that generates actions according to rules: A t =RuleEngine(D h ,Q t ) Among them, A t represents the action of round t, RuleEngine represents the rule-based dialogue engine, and D h Represents the encoded features of the conversation history, Q t Indicates the content of the conversation in round t; The preliminary model of supervised learning to achieve dialogue management is expressed as: P(A t |D h ,Q t )=Softmax(W·[D h ;Q t ]) Among them, P(A t |D h ,Q t ) represents the predicted action probability distribution, W represents the model parameters, [D h ;Q t ] represents the connection between the encoding features of the conversation history and the current conversation content; The self-learning algorithm updates the dialogue management model through a reinforcement learning framework: The active planning of dialogue content and improvement of multi-round coherence and controllability are expressed as follows: Q t+1 =Planner(D h ,Q t ,A t ) Among them, Q t+1 represents the content of the conversation in round t+1, Planner represents the active planning strategy, and D h Represents the encoded features of the conversation history, Q t Indicates the content of the conversation in round t, A t Represents the action in round t.

2. The method for rapidly constructing a financial question-answering algorithm according to claim 1, characterized in that: Described S1 comprises the following steps: S11: Using automated data collection methods based on open source datasets, web crawler technology, and log collection, we can quickly acquire large-scale question-answering data from multiple financial-related data sources, including financial news websites, reports, documents, and user questions. S12: Preprocess the collected large-scale question-answering data using natural language processing and text cleaning, including denoising, text normalization, and semantic parsing; S13: Extract features of pre-processed data using LoRa technology.

3. The method for rapidly constructing a financial question-answering algorithm according to claim 1, characterized in that: The formula for the bidirectional information decoupled multi-round dialogue network combined with reinforcement learning as a universal dialogue encoder described in S21 is as follows: H t =BIDirectionalRNN(D h ,H t-1 ,Q t ) Where J(θ) represents the objective function, θ represents the model parameters, Q represents the sequence of dialogue history, including the dialogue content of multiple rounds; A represents the action sequence, which represents the encoding selection of the bidirectional information decoupling network in each round; R t represents the reward sequence, which represents the reinforcement learning reward in each round; BIDirectionalRNN represents a bidirectional recurrent neural network for encoding the dialogue history; D h Represents the parameters of the network, which are used to control the encoding process of the bidirectional information decoupling network.

4. The method for rapidly constructing a financial question-answering algorithm according to claim 1, characterized in that: The autoencoder network described in S21 uses multi-task learning for incremental pre-training; specifically: The training objective of the autoencoder network can be expressed as: Among them, L autoencoder represents the training loss of the autoencoder network, X t represents the original input, represents the input for reconstruction; The goal of multi-task learning can be expressed as: Among them, L MTL represents the total loss of multi-task learning, α i Represents the weight of task i, Loss i That is the loss of task i; The autoregressive network is based on the Tranformer-XL structure and supports long text language model modeling.

5. The method for rapidly constructing a financial question-answering algorithm according to claim 1, characterized in that: The general conversation encoder in the general domain large-scale natural language processing model described in S22 can be expressed as: Z=Encoder(Q) Among them, Z represents the encoded features of the dialogue history, Encoder represents the universal dialogue encoder, and Q represents the sequence of dialogue history.

6. The method for rapidly constructing a financial question-answering algorithm according to claim 1, characterized in that: Described S3 comprises the following steps: S31: Constructing a financial knowledge graph; S32: Build an application named entity recognition model; S33: Applying financial knowledge graphs and entity recognition to the training and reasoning of artificial intelligence conversational interaction algorithm models.

7. A method for rapidly constructing a financial question-answering algorithm according to claim 6, characterized in that: The S31 includes: The financial knowledge graph includes the definition of entity nodes and relationship edges, where the entity set E represents various entities in the financial field, including companies, indicators, and terms. The relationship set R represents the relationship between entities, including upstream and downstream relationships and similar relationships. The adjacency matrix A is used to represent the relationship of the graph. Specifically: Among them A ij Indicates whether there is a relationship between entity i and entity j; The S32 includes: In the process of named entity recognition, the conditional random field CRF model is used to identify entities; given a text sequence T, a state sequence S is defined, where S i Indicates whether position i is a named entity; the NER module algorithm formula is as follows: Among them, Z(T) is the normalization factor, n is the length of the text sequence, k is the number of feature functions, and λ j is the feature weight, f j (i, T, S) is the jth feature function, which represents the relationship between the feature at position i and the state S; The S33 includes: Incorporating professional knowledge into AI conversational interaction algorithm models to improve the system's understanding and accuracy of financial questions. This includes: combining entity recognition results with knowledge graph information to match and reason between questions and answers; using knowledge graphs and entity recognition results to enhance the understanding capabilities of AI conversational interaction algorithm models; calculating the correlation between entities in the question and entities in the knowledge graph, and then matching questions and answers based on this correlation. Correlation (e i ,e j )=Similarity(e i ,e j )×weight(e j ) Among them, e i and e j Represent the entities in the question and the entities in the knowledge graph respectively, and the similarity (e i ,e j ) represents the similarity between entities, and the weight (e j ) represents the weight of the entity in the knowledge graph; then the entity with the highest correlation is selected for matching and reasoning between questions and answers.

8. A lifecycle management platform for financial question-answering algorithms, characterized by: The platform is used to manage the financial question-answering algorithm described in claims 1-7, and the management includes the generation, testing, deployment, monitoring and updating of the algorithm model; the lifecycle management platform includes: Graph databases facilitate users to store model parameters, dataset information, and entity relationships, enabling efficient query and retrieval capabilities; A visual operation and configuration interface for configuring algorithm model hyperparameters, selecting data sets, and setting monitoring indicators; The automated model update and version control mechanism is used to regularly detect changes in the data set, monitor the performance of the algorithm model, and automatically store the model for retraining and deployment based on the configuration parameters, thereby ensuring that the algorithm model is in the optimal state.

Images