CN113239679A - Bond position automatic detection technology based on deep learning - Google Patents

Abstract

Translated fromChinese

一种基于深度学习的债券标位自动检测技术，它涉及金融AI技术领域，它包括了以下具体模块：通用接口模块，接收聊天系统提供的信息，进行简单加工处理处理；文本解析识别模块，识别用户“投标、改标、撤标、新增”等意图，并且把相应的要素提取；重整化模块，基于用户提供的信息和历史数据库中的记录，整理成标准的数据形式；执行模块，执行具体的“投标、改标、撤标”等操作；存储器模块，存储用户的投标信息，本发明有益效果为：实现了自动检测聊天中的“投标/改标/撤标/增投”等用户意图，并通过程式实现相关的具体操作；降低了人力成本，且采用“意图识别+要素提取”的融合模型，并且融入金融领域的“经验特征”信息，目前整套系统的识别、解析准确率达到95％以上的水平，基本达到商用标准。

An automatic detection technology of bond marks based on deep learning, which relates to the field of financial AI technology, and includes the following specific modules: a general interface module, which receives information provided by a chat system and performs simple processing; a text parsing and recognition module, which recognizes The user's intention to "bid, change, withdraw, add" and other intentions, and extract the corresponding elements; the renormalization module, based on the information provided by the user and the records in the historical database, organizes it into a standard data form; the execution module, Execute specific operations such as "bidding, changing bids, withdrawing bids"; the memory module stores the bidding information of users, and the beneficial effects of the present invention are: realizing automatic detection of "bidding/modifying/cancelling/adding bids" in chatting, etc. User intentions, and related specific operations are realized through programs; labor costs are reduced, and the fusion model of "intent recognition + element extraction" is adopted, and the information of "experience characteristics" in the financial field is integrated. At present, the recognition and analysis accuracy of the entire system is Reaching the level of more than 95%, basically reaching the commercial standard.

Description

Bond position automatic detection technology based on deep learning

Technical Field

The invention relates to the technical field of financial AI (Artificial intelligence), in particular to a bond position automatic detection technology based on deep learning.

Background

In bond investment, traders of securities traders or distributors usually communicate with investors through an instant chat tool IM to determine the operation of bidding, bidding modification or bid withdrawal of a bond by the investors.

The existing processing methods mainly include two methods, one is to manually arrange the bidding data of investors according to chat contents by manual combing, adopting excel and other office software, or is a semi-manual method, namely, a rule method is adopted, rules for detecting the bidding position are defined by expert combing, whether the bidding information is the bidding information or not and effective information contained in the bidding information are judged by some keywords and some rules, with the rise of deep learning technology, a deep learning algorithm for text classification and information extraction is gradually applied to text and digital processing in the financial field, and due to the difference of particularity and language expression in the private field, a technology which can be commercialized in a general field cannot achieve a good effect in the private field of finance.

In the prior art, firstly, a manual method or a rule method is adopted, a large number of professionals are required to carry out rule combing due to various rules, and a large amount of affair repeated labor is needed in the process, so that the method is a method which is labor-consuming and labor-consuming; however, it is difficult to constrain the linguistic expression of the user, and secondly, it is a feasible direction from the technical aspect by the emerging deep learning algorithm technology, the key technology of deep learning involved in the mark position detection can be decomposed into a series of subtasks such as text classification, intention identification, information extraction and the like, such as classification, intention identification having fasttext, textcnn or a classifier based on Bert, and the information extraction having various NER models, BiLSTM + CRF, Bert + LSTM + CRF, and the latest models developed for chinese such as Graph4CNER and the like, and the models of text classification, parsing and the like have good effects in the general domain, but have specificity for the financial domain, such as: multiple ambiguity problems can be distinguished only by professional knowledge; the financial field has high requirements for accuracy effect, and the existing scheme is difficult to achieve commercial effect.

Disclosure of Invention

The invention aims to provide a deep learning-based automatic detection technology for bond positions, aiming at the defects and shortcomings in the prior art, so that the purposes of automatically detecting user intentions such as 'bidding/bid alteration/bid withdrawal/increased delivery' in chatting are realized, and related specific operations are realized through a program; the human cost is reduced, a fusion model of intention identification and element extraction is adopted, and the experience characteristic information in the financial field is fused, so that the identification and analysis accuracy of the whole set of system reaches the level of more than 95% at present and basically reaches the commercial standard.

In order to achieve the purpose, the invention adopts the following technical scheme: a bond position automatic detection technology based on deep learning comprises the following specific modules: the universal interface module is used for receiving the information provided by the chat system and carrying out simple processing treatment; the text analysis and identification module is used for identifying intentions of the user such as 'bidding, bid altering, bid withdrawing, newly adding' and the like and extracting corresponding elements; the reforming module is used for sorting the information provided by the user and the records in the historical database into a standard data form; the execution module executes specific operations of bidding, bid altering, bid removing and the like; and the memory module stores the bidding information of the user.

More optimized: the universal interface module is used for receiving the dialogue information provided by the chat system, processing the dialogue information, providing a standard data input interface, receiving the text information, the sender, the receiver, the dialogue time and other information in the chat, reading the historical transaction record information of the user in the database according to the sender information, sorting the data, packaging and transmitting the data to the analysis and identification module.

More optimized: the text analysis and identification module: the method comprises the steps of identifying the intention of a user and extracting element information, wherein the core function is to judge the intention of an investor according to conversation linguistic data of chatting, the intentions mainly comprise four types, namely bidding, bid changing, bid withdrawing and bid adding, and secondly identify the elements provided by the user, namely fifteen element information related to bonds, such as bond names, bond codes, bid positions, scalar quantities, subject/bond item rating and the like, and the specific technical scheme of a text analysis and identification module adopts a combined model scheme, namely an intention identification scheme and an element extraction scheme, and a deep neural network algorithm is realized in one model.

More optimized: the reforming module: the data is arranged into a unified and executable module which is a set of logic library, completion information is arranged into a standard form according to the intention of a user and provided element information, when the intention of the user is 'bid/newly increased/withdrawn', the data is arranged into the forms of 'bond name, mark position and scalar', when the intention of the user is 'withdrawn', the data is arranged into the forms of 'bond name, mark position, scalar' withdrawn ', the data is arranged into the forms of' bond name, mark position and scalar ', when the intention is withdrawn', the data is arranged into the forms of 'bond name, mark position, scalar and withdrawn', and the detailed process of the rearrangement is as follows: grouping the intention of the message, identifying the intention of the whole sentence message by using a rule, and setting a main key: bond name and bond code, and secondary key: the editing mode and the time limit are used as entity list splitting keys to reasonably split and combine the entity list; then iteratively supplementing necessary entities which are possibly lost to the grouped entity list, firstly, carrying out prefix tree matching by using bond keywords and a bond type library, and secondly, mapping and supplementing the key entities such as the lost bond names, bond types and bond keywords and the like by associating the front and back relations; finally, the 'label changing' logic is distinguished, and the 'before change' and 'after change' entity lists are split.

More optimized: the execution module: the method comprises the steps of executing specific operations such as bidding, bid changing, bid removing and the like, executing specific operations such as bidding, bid changing and bid removing, directly linking a database when the intention is 'bidding', writing data into the database, dividing the database into a plurality of modes when the intention is 'bid changing', wherein the modes comprise four modification modes of 'more than one modification mode, one more than one modification mode and more than one modification mode', deleting data in the database, writing the modified data into the database, and deleting the data in the database and writing a new result into the database when the intention is 'increased investment'.

More optimized: the data storage module is used for storing order data of users, and includes but is not limited to SQL, Redis and other databases.

The working principle of the invention is as follows: the universal interface module receives information transmitted by the chat system, including text information of investment of senders, receivers and bonds, and the like, performs simple processing, and transmits the information to the text analysis and identification module, and the identification module judges the intention of a user on bidding, bid changing, bid withdrawing or newly increasing according to the text information and identifies specific information elements such as bond names, bid positions, scalar quantities, time limits and the like. Then the data are transmitted into a reforming module which is arranged into different expression forms based on different intents, wherein interface packaging, intention identification, element extraction, reforming, an execution module and a fused whole set of automatic position marking detection framework are carried out; a multitask joint model algorithm built again based on the classical model; and the data reforming scheme under different operation conditions of the optimization scheme bond with the 'empirical characteristics' fused into the deep learning model is a main mode method.

After the technical scheme is adopted, the invention has the beneficial effects that: the user intentions such as 'bid/bid-changing/bid-withdrawing/bid-increasing' in the chat are automatically detected, and the related specific operation is realized through a program; the human cost is reduced, a fusion model of intention identification and element extraction is adopted, and the experience characteristic information in the financial field is fused, so that the identification and analysis accuracy of the whole set of system reaches the level of more than 95% at present and basically reaches the commercial standard.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.

FIG. 1 is a business process diagram of the present invention.

Fig. 2 is a network architecture diagram of a text parsing and recognition module.

FIG. 3 is a schematic diagram of the reforming logic of the reforming module.

Detailed Description

Referring to fig. 1 to fig. 3, the technical solution adopted by the present embodiment is: the system comprises the following specific modules: the universal interface module is used for receiving the information provided by the chat system and carrying out simple processing treatment; the text analysis and identification module is used for identifying intentions of the user such as 'bidding, bid altering, bid withdrawing, newly adding' and the like and extracting corresponding elements; the reforming module is used for sorting the information provided by the user and the records in the historical database into a standard data form; the execution module executes specific operations of bidding, bid altering, bid removing and the like; and the memory module stores the bidding information of the user.

More optimized: the universal interface module is used for receiving the dialogue information provided by the chat system, processing the dialogue information, providing a standard data input interface, receiving the text information, the sender, the receiver, the dialogue time and other information in the chat, reading the historical transaction record information of the user in the database according to the sender information, sorting the data, packaging and transmitting the data to the analysis and identification module.

More optimized: the text analysis and identification module: the method comprises the steps of identifying the intention of a user and extracting element information, wherein the core function is to judge the intention of an investor according to conversation linguistic data of chatting, the intentions mainly comprise four types, namely bidding, bid changing, bid withdrawing and bid adding, and secondly identify the elements provided by the user, namely fifteen element information related to bonds, such as bond names, bond codes, bid positions, scalar quantities, subject/bond item rating and the like, and the specific technical scheme of a text analysis and identification module adopts a combined model scheme, namely an intention identification scheme and an element extraction scheme, and a deep neural network algorithm is realized in one model.

More optimized: the reforming module: the data is arranged into a unified and executable module which is a set of logic library, completion information is arranged into a standard form according to the intention of a user and provided element information, when the intention of the user is 'bid/newly increased/withdrawn', the data is arranged into the forms of 'bond name, mark position and scalar', when the intention of the user is 'withdrawn', the data is arranged into the forms of 'bond name, mark position, scalar' withdrawn ', the data is arranged into the forms of' bond name, mark position and scalar ', when the intention is withdrawn', the data is arranged into the forms of 'bond name, mark position, scalar and withdrawn', and the detailed process of the rearrangement is as follows: grouping the intention of the message, identifying the intention of the whole sentence message by using a rule, and setting a main key: bond name and bond code, and secondary key: the editing mode and the time limit are used as entity list splitting keys to reasonably split and combine the entity list; then iteratively supplementing necessary entities which are possibly lost to the grouped entity list, firstly, carrying out prefix tree matching by using bond keywords and a bond type library, and secondly, mapping and supplementing the key entities such as the lost bond names, bond types and bond keywords and the like by associating the front and back relations; finally, the 'label changing' logic is distinguished, and the 'before change' and 'after change' entity lists are split.

More optimized: the execution module: the method comprises the steps of executing specific operations such as bidding, bid changing, bid removing and the like, executing specific operations such as bidding, bid changing and bid removing, directly linking a database when the intention is 'bidding', writing data into the database, dividing the database into a plurality of modes when the intention is 'bid changing', wherein the modes comprise four modification modes of 'more than one modification mode, one more than one modification mode and more than one modification mode', deleting data in the database, writing the modified data into the database, and deleting the data in the database and writing a new result into the database when the intention is 'increased investment'.

More optimized: the data storage module is used for storing order data of users, and includes but is not limited to SQL, Redis and other databases.

The working principle of the invention is as follows: the universal interface module receives information transmitted by the chat system, including text information of investment of senders, receivers and bonds, and the like, performs simple processing, and transmits the information to the text analysis and identification module, and the identification module judges the intention of a user on bidding, bid changing, bid withdrawing or newly increasing according to the text information and identifies specific information elements such as bond names, bid positions, scalar quantities, time limits and the like. Then the data are transmitted into a reforming module which is arranged into different expression forms based on different intents, wherein interface packaging, intention identification, element extraction, reforming, an execution module and a fused whole set of automatic position marking detection framework are carried out; a multitask joint model algorithm built again based on the classical model; and the data reforming scheme under different operation conditions of the optimization scheme bond with the 'empirical characteristics' fused into the deep learning model is a main mode method.

The above description is only for the purpose of illustrating the technical solutions of the present invention and not for the purpose of limiting the same, and other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the scope of the claims of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims (8)

1. The utility model provides a bond position automatic check out technique based on deep learning which characterized in that: the system comprises the following specific modules:

a) the universal interface module is used for receiving the information provided by the chat system and carrying out simple processing treatment;

b) the text analysis and identification module is used for identifying intentions of the user such as 'bidding, bid altering, bid withdrawing, newly adding' and the like and extracting corresponding elements;

c) the reforming module is used for sorting the information provided by the user and the records in the historical database into a standard data form;

d) the execution module executes specific operations of bidding, bid altering, bid removing and the like;

e) and the memory module stores the bidding information of the user.

2. The deep learning based bond position automatic detection technology according to claim 1, characterized in that: the universal interface module is used for receiving the dialogue information provided by the chat system, processing the dialogue information, providing a standard data input interface, receiving the text information, the sender, the receiver, the dialogue time and other information in the chat, reading the historical transaction record information of the user in the database according to the sender information, sorting the data, packaging and transmitting the data to the analysis and identification module.

3. The deep learning based bond position automatic detection technology according to claim 1, characterized in that: the text analysis and identification module: the method comprises the steps of identifying the intention of a user and extracting element information, wherein the core function is to judge the intention of an investor according to conversation linguistic data of chatting, and the intention mainly comprises four types of element information, namely bidding, bid changing, bid withdrawing and bid adding, and secondly identifies the elements provided by the user, namely fifteen element information related to bonds, such as bond names, bond codes, mark positions, scalar quantities, subject/bond rating and the like.

4. The deep learning based bond position automatic detection technology according to claim 3, characterized in that: the specific technical scheme of the text analysis and identification module adopts a combined model scheme, namely a deep neural network algorithm realized in a model by 'intention identification' and 'element extraction'.

5. The deep learning based bond position automatic detection technology according to claim 1, characterized in that: the reforming module: the data is arranged into a unified and executable module which is a set of logic library, completion information is arranged into a standard form according to the intention of a user and provided element information, when the intention of the user is ' bid/newly increased/withdrawn ', the data is arranged into the forms of ' bond name, mark position and scalar ', when the intention of the user is ' withdrawn ', the data is arranged into the forms of ' bond name, mark position and scalar ', when the intention is withdrawn, the data is arranged into the forms of ' bond name, mark position, scalar ' and withdrawn '.

6. The deep learning based bond position automatic detection technology according to claim 5, characterized in that: the specific process of the reforming comprises the following steps: grouping the intention of the message, identifying the intention of the whole sentence message by using a rule, and setting a main key: bond name and bond code, and secondary key: the editing mode and the time limit are used as entity list splitting keys to reasonably split and combine the entity list; then iteratively supplementing necessary entities which are possibly lost to the grouped entity list, firstly, carrying out prefix tree matching by using bond keywords and a bond type library, and secondly, mapping and supplementing the key entities such as the lost bond names, bond types and bond keywords and the like by associating the front and back relations; finally, the 'label changing' logic is distinguished, and the 'before change' and 'after change' entity lists are split.

7. The deep learning based bond position automatic detection technology according to claim 1, characterized in that: the execution module: the method comprises the steps of executing specific operations such as bidding, bid changing, bid removing and the like, executing specific operations such as bidding, bid changing and bid removing, directly linking a database when the intention is 'bidding', writing data into the database, dividing the database into a plurality of modes when the intention is 'bid changing', wherein the modes comprise four modification modes of 'more than one modification mode, one more than one modification mode and more than one modification mode', deleting data in the database, writing the modified data into the database, and deleting the data in the database and writing a new result into the database when the intention is 'increased investment'.

8. The deep learning based bond position automatic detection technology according to claim 1, characterized in that: the data storage module is used for storing order data of users, and includes but is not limited to SQL, Redis and other databases.

Images