A method, system and device for generating inquiry letter
By applying machine learning to historical inquiry letters and prospectuses, a target generation model is established to generate a rule dictionary and automatically generate inquiry letters, solving the problem of manpower and time consumption in writing inquiry letters and improving the efficiency of inquiry letter generation.
Patent Information
- Application Number
- CN202111277389.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-29
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2041-10-29
AI Technical Summary
In the current technology, the writing of inquiry letters requires purely manual operation, which is labor-intensive, time-consuming and costly.
By applying machine learning to historical inquiry letters and prospectuses, a target generation model is established to generate a rule dictionary, which is used to automatically generate inquiry letters, reducing the need for manual drafting.
It has enabled the automated generation of inquiry letters, saving manpower and time costs and improving the efficiency of inquiry letter generation.
Smart Images

Figure CN113901778B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present application relate to the field of algorithm models, and in particular, to a method, system and device for generating an inquiry letter. Background Art
[0002] The registration system is a securities issuance method with information disclosure as its core. Regarding the authenticity, completeness and accuracy of information disclosure in public offerings, mergers and acquisitions, restructuring and re-issuances, the exchange will send an inquiry letter to the issuer, and the issuer and issuing intermediaries such as sponsors, accounting firms and law firms will answer. The authenticity, completeness and accuracy of information disclosure will be supplemented through multiple rounds of inquiries, so that investors can fully understand the investment value and risks of the issuer, reduce capital market risks, and enhance the capital market's ability to serve the real economy.
[0003] Sending an inquiry letter to the issuer is an important way to exercise the right to review the issuance of securities under the registration system. The inquiry letter raises inquiries about the abnormal review points in the issuer's prospectus, and the issuer responds to the inquiry letter, thereby ensuring that the issuer's information is fully disclosed and protecting the investors' right to know. However, writing an inquiry letter requires the auditor to read the application materials, review relevant information, determine whether the information disclosed in the prospectus is authentic, complete and accurate, and write a report to inquire the issuer, sponsor, accounting firm or law firm.
[0004] However, in the actual work process, the writing of inquiry letters needs to be purely manual, which is labor-intensive and manpower-consuming. Summary of the invention
[0005] The embodiments of the present application provide a method, system and device for generating an inquiry letter, which are used to generate an inquiry letter for abnormal review points in the issuer's prospectus, assist in manual writing, and save manpower and time costs.
[0006] A method for generating an inquiry letter provided in an embodiment of the present application includes:
[0007] Obtaining a target generation model, wherein the target generation model is obtained by performing machine learning training on an initial generation model using historical inquiry letters and historical prospectuses, wherein a rule dictionary for generating a target inquiry letter is stored in the target generation model, wherein the historical inquiry letters and the historical prospectuses have a corresponding relationship, and the historical inquiry letters are generated based on the historical prospectuses, and wherein the rule dictionary is obtained by performing machine learning on the historical inquiry letters and the historical prospectuses in combination with manual experience;
[0008] Get the text vector of the target prospectus;
[0009] The text vector of the target prospectus is input into the target generation model, and the target generation model outputs a target inquiry letter corresponding to the target prospectus according to the rule dictionary.
[0010] Optionally, before acquiring the target generation model, the method further includes:
[0011] Obtaining the historical inquiry letter and the historical prospectus, and using a text vector algorithm to obtain text vectors of the historical inquiry letter and the historical prospectus;
[0012] The text vectors of the historical inquiry letters and the historical prospectuses are input into the initial prediction model as training samples, and the initial generation model is subjected to machine learning training using the training samples to obtain the target generation model, in which the rule dictionary is stored.
[0013] Optionally, the rule dictionary is obtained by machine learning of historical inquiry letters and historical prospectuses and combined with manual experience and includes:
[0014] Using a classification algorithm to decompose the historical inquiry letter into inquiry background paragraphs and inquiry question paragraphs, and identifying the subdivided questions contained in the inquiry question paragraphs, wherein one audit point corresponds to at least one inquiry background paragraph and at least one inquiry question paragraph, and the inquiry background paragraph and the inquiry question paragraph corresponding to the same audit point have an associated relationship;
[0015] Determine the audit points corresponding to the subdivided questions, wherein one audit point corresponds to at least one subdivided question, and form a first directory indicating the correspondence between the audit points and the subdivided questions;
[0016] Obtaining keywords contained in the segmented questions, and using the keywords as the inquiry direction of the segmented questions;
[0017] The subdivided questions corresponding to each of the audit points are clustered according to the inquiry direction to obtain a second directory representing the correspondence between the audit points, the inquiry directions and the subdivided questions, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one subdivided question;
[0018] Input the subdivided questions into a neural network algorithm model to generate question templates, and obtain a third directory representing the correspondence between the audit points, the inquiry directions, and the question templates, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one question template;
[0019] Determine the triggering condition of the audit point according to the subdivided question, and obtain a fourth directory representing the corresponding relationship between the audit point, the inquiry direction, the question template and the triggering condition, wherein the triggering condition is used to determine whether to inquire about the audit point;
[0020] Performing machine learning on the inquiry background paragraph and the historical prospectus to obtain a summary rewriting rule for generating the inquiry background paragraph;
[0021] The rule dictionary includes the fourth directory and the summary rewrite rule.
[0022] Optionally, obtaining the keywords contained in the segmented question and using the keywords as the query direction of the segmented question includes:
[0023] Performing word segmentation matching based on word strength on the segmented problem to obtain a preprocessed word segmentation set, and using a text vector algorithm and a text classification algorithm to process the preprocessed word segmentation set to obtain a keyword seed word library;
[0024] Expanding the keyword seed word library with a synonym dictionary based on artificial experience to obtain a keyword word library;
[0025] or using a text vector algorithm to calculate the similarity of words in the keyword seed word library to expand the keyword seed word library to obtain the keyword word library;
[0026] The keywords included in the segmented question are obtained according to the keyword database, and the keywords are used as the query direction of the segmented question.
[0027] Optionally, inputting the segmented problem into a neural network algorithm model to generate a problem template includes:
[0028] Using a text vector algorithm to obtain a text vector for the segmented problem, inputting the text vector for the segmented problem as a training sample into an initial text recognition model for machine learning training, and obtaining a target text recognition model, wherein the target text recognition model stores a recognition method for entity text, attribute text, and attribute value text in the segmented problem;
[0029] The entity text, the attribute text and the attribute value text in the segmented question are replaced with placeholders to obtain the question template.
[0030] Optionally, the triggering condition for determining the audit key points according to the segmented questions includes:
[0031] The triggering conditions of the audit points are divided into two types: rule mode and content mode. The triggering conditions of the rule mode include:
[0032] Presetting the attribute value threshold according to the attribute value of the detailed question, and determining to trigger the audit key point if the attribute value in the historical prospectus does not meet the attribute value threshold, wherein the attribute value in the detailed question in the historical inquiry letter corresponds to the attribute value of the corresponding audit key point in the historical prospectus;
[0033] The triggering conditions of the content mode include:
[0034] A binary classification model for determining whether there are abnormal audit points is obtained by performing machine learning training using the historical inquiry letters and the historical prospectuses as training samples;
[0035] When the target prospectus is input into the binary classification model and a conclusion is drawn that abnormal audit points exist, it is determined that the audit points are triggered.
[0036] Optionally, the performing machine learning on the inquiry background paragraph and the historical prospectus to obtain a summary rewriting rule for generating the inquiry background paragraph includes:
[0037] Perform machine learning training using the inquiry background paragraph and the historical prospectus as training samples to obtain a corresponding relationship between the inquiry background paragraph and the content in the historical prospectus;
[0038] A summary rewriting rule for generating the inquiry background paragraph is determined according to the corresponding relationship, and the summary rewriting rule is used to summarize the text paragraphs in the historical prospectus that have a corresponding relationship with the inquiry background paragraph to generate the inquiry background paragraph.
[0039] An inquiry letter generation system provided in an embodiment of the present application includes:
[0040] an acquisition unit, configured to acquire a target generation model, wherein the target generation model is obtained by performing machine learning training on an initial generation model using historical inquiry letters and historical prospectuses, wherein a rule dictionary for generating a target inquiry letter is stored in the target generation model, wherein the historical inquiry letters have a corresponding relationship with the historical prospectuses, and the historical inquiry letters are generated based on the historical prospectuses, and wherein the rule dictionary is obtained by performing machine learning on the historical inquiry letters and the historical prospectuses in combination with manual experience;
[0041] The acquisition unit is further used to acquire the text vector of the target prospectus;
[0042] An input unit is used to input the text vector of the target prospectus into the target generation model, and the target generation model outputs a target inquiry letter corresponding to the target prospectus according to the rule dictionary.
[0043] Optionally, the acquisition unit is further configured to acquire the historical inquiry letter and the historical prospectus before acquiring the target generation model, and to acquire text vectors of the historical inquiry letter and the historical prospectus using a text vector algorithm;
[0044] The text vectors of the historical inquiry letters and the historical prospectuses are input into the initial prediction model as training samples, and the initial generation model is subjected to machine learning training using the training samples to obtain the target generation model, in which the rule dictionary is stored.
[0045] Optionally, the acquisition unit is specifically used to use a classification algorithm to decompose the historical inquiry letter into an inquiry background paragraph and an inquiry question paragraph, and identify the subdivided questions contained in the inquiry question paragraph, wherein one audit point corresponds to at least one inquiry background paragraph and at least one inquiry question paragraph, and the inquiry background paragraph and the inquiry question paragraph corresponding to the same audit point have an associated relationship;
[0046] Confirming the audit points corresponding to the subdivided questions, wherein one audit point corresponds to at least one subdivided question, and forming a first directory indicating the correspondence between the audit points and the subdivided questions;
[0047] Obtaining keywords contained in the segmented questions, and using the keywords as the inquiry direction of the segmented questions;
[0048] The subdivided questions corresponding to each of the audit points are clustered according to the inquiry direction to obtain a second directory representing the correspondence between the audit points, the inquiry directions and the subdivided questions, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one subdivided question;
[0049] Input the subdivided questions into a neural network algorithm model to generate question templates, and obtain a third directory representing the correspondence between the audit points, the inquiry directions, and the question templates, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one question template;
[0050] Determine the triggering condition of the audit point according to the subdivided question, and obtain a fourth directory representing the corresponding relationship between the audit point, the inquiry direction, the question template and the triggering condition, wherein the triggering condition is used to determine whether to inquire about the audit point;
[0051] Performing machine learning on the inquiry background paragraph and the historical prospectus to obtain a summary rewriting rule for generating the inquiry background paragraph;
[0052] The rule dictionary includes the fourth directory and the summary rewrite rule.
[0053] Optionally, the acquisition unit is further configured to perform word segmentation matching based on the strength of words on the segmented questions to obtain a preprocessed word segmentation set, and use a text vector algorithm and a text classification algorithm to process the preprocessed word segmentation set to obtain a keyword seed word library;
[0054] Expanding the keyword seed word library with a synonym dictionary based on artificial experience to obtain a keyword word library;
[0055] or using a text vector algorithm to calculate the similarity of words in the keyword seed word library to expand the keyword seed word library to obtain the keyword word library;
[0056] The keywords included in the segmented question are obtained according to the keyword database, and the keywords are used as the query direction of the segmented question.
[0057] Optionally, the acquisition unit is further used to use a text vector algorithm to acquire the text vector of the segmented problem, and input the text vector of the segmented problem as a training sample into an initial text recognition model for machine learning training to obtain a target text recognition model, wherein the target text recognition model stores a recognition method of entity text, attribute text, and attribute value text in the segmented problem;
[0058] The entity text, the attribute text and the attribute value text in the segmented question are replaced with placeholders to obtain the question template.
[0059] Optionally, the acquisition unit is further configured to determine a trigger condition for the audit key points according to the segmented questions. The trigger condition for the audit key points is divided into two types: a rule mode and a content mode. The trigger condition for the rule mode includes:
[0060] Presetting the attribute value threshold according to the attribute value of the detailed question, and determining to trigger the audit key point if the attribute value in the historical prospectus does not meet the attribute value threshold, wherein the attribute value in the detailed question in the historical inquiry letter corresponds to the attribute value of the corresponding audit key point in the historical prospectus;
[0061] The triggering conditions of the content mode include:
[0062] A binary classification model for determining whether there are abnormal audit points is obtained by performing machine learning training using the historical inquiry letters and the historical prospectuses as training samples;
[0063] When the target prospectus is input into the binary classification model and a conclusion is drawn that abnormal audit points exist, it is determined that the audit points are triggered.
[0064] Optionally, the acquisition unit is further configured to use the inquiry background paragraph and the historical prospectus as training samples to perform machine learning training to obtain a correspondence between the inquiry background paragraph and the content in the historical prospectus;
[0065] A summary rewriting rule for generating the inquiry background paragraph is determined according to the corresponding relationship, and the summary rewriting rule is used to summarize the text paragraphs in the historical prospectus that have a corresponding relationship with the inquiry background paragraph to generate the inquiry background paragraph.
[0066] The present application also provides an inquiry letter generating device including:
[0067] CPU, memory and input / output interface;
[0068] The memory is a short-term storage memory or a persistent storage memory;
[0069] The central processing unit is configured to communicate with the memory and execute instruction operations in the memory to perform the aforementioned inquiry letter generation method.
[0070] An embodiment of the present application also provides a computer-readable storage medium, comprising instructions, which, when executed on a computer, enable the computer to execute the aforementioned inquiry letter generating method.
[0071] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:
[0072] A target generation model for generating a target inquiry letter is obtained by performing machine learning on historical inquiry letters and historical prospectuses. The target generation model stores a rule dictionary for generating a target inquiry letter, so that when the target prospectus is input into the target generation model, the target generation model can output a target inquiry letter, reducing the manpower and time costs required for pure manual writing. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1 A schematic diagram of an implementation method of the inquiry letter generation method provided in an embodiment of the present application;
[0074] Figure 2 A schematic diagram of another implementation of the method for generating an inquiry letter provided in an embodiment of the present application;
[0075] Figure 3 A schematic diagram of an implementation of the inquiry letter generation system provided in an embodiment of the present application;
[0076] Figure 4 A schematic diagram of an implementation of an inquiry letter generating device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0077] The embodiments of the present application provide a method, system and device for generating an inquiry letter, which are used to generate an inquiry letter for abnormal review points in the issuer's prospectus, assist in manual writing, and save manpower and time.
[0078] The registration system is a securities issuance method with information disclosure as its core. Regarding the authenticity, completeness and accuracy of information disclosure in public offerings, mergers and acquisitions, restructuring and re-issuances, the exchange will send an inquiry letter to the issuer, and the issuer and issuing intermediaries such as sponsors, accounting firms and law firms will answer. The authenticity, completeness and accuracy of information disclosure will be supplemented through multiple rounds of inquiries, so that investors can fully understand the investment value and risks of the issuer, reduce capital market risks, and enhance the capital market's ability to serve the real economy.
[0079] Sending an inquiry letter to the issuer is an important way to exercise the right to review the issuance of securities under the registration system. The inquiry letter raises inquiries about the abnormal review points in the issuer's prospectus, and the issuer responds to the inquiry letter, thereby ensuring that the issuer's information is fully disclosed and protecting the investors' right to know. However, writing an inquiry letter requires the auditor to read the application materials, review relevant information, determine whether the information disclosed in the prospectus is authentic, complete and accurate, and write a report to inquire the issuer, sponsor, accounting firm or law firm.
[0080] However, in the actual work process, the writing of the inquiry letter needs to be purely manual, which is labor-intensive and time-consuming. Therefore, a method that can use a computer to automatically generate an inquiry letter is needed to assist in the manual writing of the inquiry letter, and the embodiment of the present application can achieve such a technical effect. In the embodiment of the present application, for the abnormal audit points that appear in the issuer's prospectus and other information disclosure documents, suitable inquiry directions and subdivided questions are recommended, and the question template is filled with entities, attributes and attribute values to generate an inquiry question paragraph, and an inquiry background paragraph is generated, and finally a relatively complete inquiry letter is generated. The generated inquiry letter only needs to be manually polished, and it can be sent to the issuer for inquiry without even manual polishing.
[0081] It should be noted that in the embodiments of the present application, the inquiry letter is used to refer to the text materials that inquire about the public text materials of the issuer, including but not limited to the form of the inquiry letter, and the prospectus is used to refer to the text materials related to the securities issuance disclosed by the issuer, including but not limited to the form of the prospectus. In order to simplify the language expression and make the technical solution clearer during the writing process, this naming method is adopted.
[0082] See also Figure 1 An implementation of a method for generating an inquiry letter provided in an embodiment of the present application includes steps 101 to 103.
[0083] 101. Obtain a target generation model.
[0084] The target generation model is established in advance and is used to generate a target inquiry letter. Since the target inquiry letter has a corresponding relationship with the target prospectus, the target inquiry letter is used to initiate inquiries to abnormal review points in the target prospectus. Therefore, the target generation model is specifically obtained by training the initial generation model through machine learning using historical inquiry letters and historical prospectuses. The target generation model stores a rule dictionary for generating the target inquiry letter, which is obtained through machine learning of historical inquiry letters and historical prospectuses combined with manual experience.
[0085] The rule dictionary includes rules for generating the inquiry question paragraph of the target inquiry letter and rules for generating the inquiry background paragraph of the target inquiry letter. In the target generation model establishment stage, the historical inquiry letter is first decomposed into the inquiry background paragraph and the inquiry question paragraph by using the classification algorithm according to the text features in the historical inquiry letter. The text features can be represented by text vectors, that is, the text of the historical inquiry letter is converted into a text vector corresponding to the semantics of the historical inquiry letter by using a text vector algorithm, such as an embedding algorithm, and the historical inquiry letter is decomposed into the inquiry background paragraph and the inquiry question paragraph according to the text vector. Other text semantic analysis methods that can achieve the same or similar effects can also be used to decompose the historical inquiry letter into the inquiry background paragraph and the inquiry question paragraph. The specifics are not limited here.
[0086] After obtaining the inquiry question paragraph, identify the sub-questions contained in the inquiry question paragraph. Similarly, a text vector algorithm, such as an embedding algorithm, can be used to further decompose the inquiry question paragraph into sub-questions, or other text semantic analysis methods that can achieve the same or similar effects can be used to further decompose the inquiry question paragraph into sub-questions, wherein one audit point corresponds to at least one inquiry background paragraph and at least one inquiry question paragraph, and the inquiry background paragraph and inquiry question paragraph corresponding to the same audit point have a correlation relationship.
[0087] Determine the audit points corresponding to the detailed questions. Since a question paragraph generally corresponds to one audit point, the detailed questions within the question paragraph generally also correspond to this audit point. One audit point corresponds to at least one detailed question, and a first directory representing the correspondence between the audit points and the detailed questions is formed.
[0088] Use the word segmentation algorithm to obtain the keywords contained in the segmented questions, and use the keywords as the query direction of the segmented questions, wherein the word segmentation algorithm can be a BPE algorithm or other algorithms that can achieve the same or similar effects, which are not limited here. Specifically, first use the word segmentation algorithm to remove non-key phrases in the segmented questions to obtain a preprocessed word segmentation set, calculate the tf-idf value of the text vector of the word segmentation in the preprocessed word segmentation set, and judge the word segmentation belonging to the keyword based on the grammatical components of the word segmentation in the sentence to obtain a keyword seed word library. Specifically, the aforementioned process of obtaining a keyword seed word library can be achieved by constructing a keyword classification model, such as using the text vector, tf-idf value and grammatical components of the word segmentation in the preprocessed word segmentation set as the input of the keyword classification model, and outputting whether the word segmentation in the preprocessed word segmentation set is a keyword, wherein the keyword classification model can select RNN, CNN or Transformer as the encoder, and superimpose a sigmoid classifier to achieve it. Since different historical inquiry letters and historical prospectuses are written by different people, they may have different wording styles, which makes the coverage of keywords in the keyword seed word library smaller, so it is necessary to expand the keyword seed word library to form a keyword word library. Specifically, the keyword word library can be obtained by expanding the keyword seed word library with a synonym dictionary based on artificial experience, or by using a text vector algorithm to calculate the similarity of words in the keyword seed word library to expand the keyword seed word library to obtain a keyword word library.
[0089] The keywords included in the subdivided questions are obtained from the keyword word library, and the keywords are used as the inquiry directions of the subdivided questions. The subdivided questions corresponding to each audit point are clustered according to the inquiry directions, and a second directory representing the correspondence between the audit points, inquiry directions and subdivided questions is obtained, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one subdivided question.
[0090] The segmented questions are input into the neural network algorithm model to generate question templates, and a third directory representing the correspondence between audit points, inquiry directions and question templates is obtained, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one question template.
[0091] Specifically, a text vector algorithm is used to obtain text vectors of segmented questions, and the text vectors of segmented questions are input into an initial text recognition model as training samples for machine learning training to obtain a target text recognition model. The target text recognition model stores the recognition methods of entity text, attribute text, and attribute value text in the segmented questions. The entity text, attribute text, and attribute value text in the segmented questions are replaced with placeholders to obtain a question template.
[0092] For example, the entities in the segmentation question mainly include date, company name, person name, amount and quantity, etc. The attribute value is the annual value, the amount value or the ratio value used to represent the entity attribute, etc. The attribute is used to represent the attribute of the entity. Since the entity is closely related to the content in the historical prospectus, the entity, attribute and attribute value in the segmentation question should be generated according to the prospectus, so a question template can be generated first. As long as the entity, attribute and attribute value are filled in the question template, an inquiry letter for the prospectus can be generated. For example, the detailed question in the historical inquiry letter is "In 2018, 2019 and 2020, the company's overseas sales revenue accounted for 43.49%, 44.71% and 23.57% of the main business income respectively. Please explain the reason for the decline in main business income", then the question template after removing entities, attributes and attribute values and filling in placeholders is "In [year], [year] and [year], the company's overseas sales revenue accounted for [POOSITMBI], [POOSITMBI] and [POOSITMBI] respectively. Please explain the reason for the decline in main business income."
[0093] Afterwards, the triggering conditions of the audit points are determined based on the segmented questions, and a fourth directory is obtained that represents the correspondence between the audit points, inquiry directions, question templates and triggering conditions. The triggering conditions are used to determine whether to inquire about the audit points.
[0094] Specifically, the triggering conditions of the audit points are divided into two types: rule mode and content mode. The triggering conditions of the rule mode include:
[0095] An attribute value threshold is preset based on the attribute value of the detailed question. If the attribute value in the historical prospectus does not meet the attribute value threshold, the audit points are determined to be triggered. The attribute value threshold can be obtained through formulas, regular expressions and four arithmetic operations. The attribute values in the detailed questions in the historical inquiry letters correspond to the attribute values of the corresponding audit points in the historical prospectuses.
[0096] The trigger conditions for content mode include:
[0097] By using historical inquiry letters and historical prospectuses as training samples for machine learning training, a binary classification model is obtained to determine whether there are abnormal audit points. When the target prospectus is input into the binary classification model and the conclusion is that there are abnormal audit points, the audit points are determined to be triggered.
[0098] Machine learning is performed on the inquiry background paragraph and historical prospectuses to obtain summary rewriting rules for generating the inquiry background paragraph.
[0099] Specifically, when the content in the prospectus meets the triggering conditions of the audit points, an inquiry background paragraph needs to be generated in the inquiry letter. Therefore, the inquiry background paragraph and the historical prospectus are used as training samples for machine learning training to obtain the corresponding relationship between the inquiry background paragraph and the content in the historical prospectus;
[0100] A summary rewriting rule for generating the inquiry background paragraph is determined based on the corresponding relationship. The summary rewriting rule is used to rewrite the summary of the text paragraphs in the historical prospectus that have a corresponding relationship with the inquiry background paragraph to generate the inquiry background paragraph.
[0101] For example, when an audit point is triggered, first identify the paragraph in the prospectus that triggers the audit point, use the inquiry direction corresponding to the audit point as the keyword of the inquiry background paragraph, and preferentially extract sentences containing keywords from the prospectus. Identify the pronouns in these sentences, replace the anaphora with the antecedent to make the sentence more fluent, thereby generating an inquiry background paragraph.
[0102] The rule dictionary is determined to include the fourth directory and summary rewriting rules, and a target inquiry letter corresponding to the target prospectus is generated according to the rule dictionary.
[0103] 102. Get the text vector of the target prospectus.
[0104] Since the input of the target generation model is a text vector, it is necessary to first use a text algorithm to obtain the text vector of the target prospectus so that the target generation model can generate a target inquiry letter corresponding to the target prospectus based on the text vector of the target prospectus.
[0105] 103. Input the text vector of the target prospectus into the target generation model, and the target generation model outputs the target inquiry letter corresponding to the target prospectus according to the rule dictionary.
[0106] After the text vector of the target prospectus is input into the target generation model, the rule dictionary is matched with the text vector of the target prospectus. When the triggering conditions of the audit points are met, the rule dictionary performs similarity matching and keyword matching based on the text vector for the paragraph in the target prospectus that triggers the audit points. If the paragraph meets a certain question template, the question template is filled with entities, attributes and attribute values to generate segmented questions. At least one segmented question constitutes the inquiry question paragraph of the target inquiry letter. The inquiry background paragraph is generated according to the summary rewriting rules. Specifically, the inquiry direction corresponding to the audit points is used as the keyword of the inquiry background paragraph. Sentences containing keywords are preferentially extracted from the prospectus, the pronouns in these sentences are identified, and the anaphora is replaced with the antecedent to make the sentence more fluent, thereby generating the inquiry background paragraph. Among them, the audit points, inquiry question paragraphs and inquiry background paragraphs have a corresponding relationship. A target inquiry letter includes at least one set of audit points, inquiry question paragraphs and inquiry background paragraphs.
[0107] In an embodiment of the present application, a target generation model for generating a target inquiry letter is obtained by performing machine learning on historical inquiry letters and historical prospectuses. The target generation model stores a rule dictionary for generating a target inquiry letter, so that when the target prospectus is input into the target generation model, the target generation model can output a target inquiry letter, thereby reducing the manpower and time costs required for pure manual writing.
[0108] See also Figure 2 Another implementation of a method for generating an inquiry letter provided in an embodiment of the present application includes steps 201 to 204.
[0109] 201. Get the text vectors of historical inquiry letters and historical prospectuses.
[0110] Since the method provided in the embodiment of the present application uses a machine learning method to establish a target generation model, the historical inquiry letter and the historical prospectus must first be used as training samples to perform machine learning training on the initial generation model, and the input of the initial generation model is the text vector of the historical inquiry letter and the historical prospectus, so it is necessary to first obtain the text vector of the historical inquiry letter and the historical prospectus. Specifically, through a text vector algorithm, such as an embedding algorithm, the text of the historical inquiry letter is converted into a text vector corresponding to the semantics of the historical inquiry letter, and the text of the historical prospectus is converted into a text vector corresponding to the semantics of the historical prospectus.
[0111] 202. The text vectors of historical inquiry letters and historical prospectuses are input into the initial prediction model as training samples, and the initial generation model is trained by machine learning using the training samples to obtain a target generation model, in which a rule dictionary is stored.
[0112] Specifically, the rule dictionary includes rules for generating inquiry question paragraphs of the target inquiry letter and rules for generating inquiry background paragraphs of the target inquiry letter. In the target generation model establishment stage, the historical inquiry letter is first decomposed into inquiry background paragraphs and inquiry question paragraphs using a classification algorithm based on the text features in the historical inquiry letter. The text features can be represented by text vectors, that is, a text vector algorithm, such as an embedding algorithm, is used to convert the historical inquiry letter text into a text vector corresponding to the semantics of the historical inquiry letter, and the historical inquiry letter is decomposed into inquiry background paragraphs and inquiry question paragraphs based on the text vector. Other text semantic analysis methods that can achieve the same or similar effects can also be used to decompose the historical inquiry letter into inquiry background paragraphs and inquiry question paragraphs, which are not specifically limited here.
[0113] After obtaining the inquiry question paragraph, identify the sub-questions contained in the inquiry question paragraph. Similarly, a text vector algorithm, such as an embedding algorithm, can be used to further decompose the inquiry question paragraph into sub-questions, or other text semantic analysis methods that can achieve the same or similar effects can be used to further decompose the inquiry question paragraph into sub-questions, wherein one audit point corresponds to at least one inquiry background paragraph and at least one inquiry question paragraph, and the inquiry background paragraph and inquiry question paragraph corresponding to the same audit point have a correlation relationship.
[0114] Determine the audit points corresponding to the detailed questions. Since a question paragraph generally corresponds to one audit point, the detailed questions within the question paragraph generally also correspond to this audit point. One audit point corresponds to at least one detailed question, and a first directory representing the correspondence between the audit points and the detailed questions is formed.
[0115] Use the word segmentation algorithm to obtain the keywords contained in the segmented questions, and use the keywords as the query direction of the segmented questions, wherein the word segmentation algorithm can be a BPE algorithm or other algorithms that can achieve the same or similar effects, which are not limited here. Specifically, first use the word segmentation algorithm to remove non-key phrases in the segmented questions to obtain a preprocessed word segmentation set, calculate the tf-idf value of the text vector of the word segmentation in the preprocessed word segmentation set, and judge the word segmentation belonging to the keyword based on the grammatical components of the word segmentation in the sentence to obtain a keyword seed word library. Specifically, the aforementioned process of obtaining a keyword seed word library can be achieved by constructing a keyword classification model, such as using the text vector, tf-idf value and grammatical components of the word segmentation in the preprocessed word segmentation set as the input of the keyword classification model, and outputting whether the word segmentation in the preprocessed word segmentation set is a keyword, wherein the keyword classification model can select RNN, CNN or Transformer as the encoder, and superimpose a sigmoid classifier to achieve it. Since different historical inquiry letters and historical prospectuses are written by different people, they may have different wording styles, which makes the coverage of keywords in the keyword seed word library smaller, so it is necessary to expand the keyword seed word library to form a keyword word library. Specifically, the keyword word library can be obtained by expanding the keyword seed word library with a synonym dictionary based on artificial experience, or by using a text vector algorithm to calculate the similarity of words in the keyword seed word library to expand the keyword seed word library to obtain a keyword word library.
[0116] The keywords included in the subdivided questions are obtained from the keyword word library, and the keywords are used as the inquiry directions of the subdivided questions. The subdivided questions corresponding to each audit point are clustered according to the inquiry directions, and a second directory representing the correspondence between the audit points, inquiry directions and subdivided questions is obtained, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one subdivided question.
[0117] The segmented questions are input into the neural network algorithm model to generate question templates, and a third directory representing the correspondence between audit points, inquiry directions and question templates is obtained, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one question template.
[0118] Specifically, a text vector algorithm is used to obtain text vectors of segmented questions, and the text vectors of segmented questions are input into an initial text recognition model as training samples for machine learning training to obtain a target text recognition model. The target text recognition model stores the recognition methods of entity text, attribute text, and attribute value text in the segmented questions. The entity text, attribute text, and attribute value text in the segmented questions are replaced with placeholders to obtain a question template.
[0119] For example, the entities in the segmentation question mainly include date, company name, person name, amount and quantity, etc. The attribute value is the annual value, the amount value or the ratio value used to represent the entity attribute, etc. The attribute is used to represent the attribute of the entity. Since the entity is closely related to the content in the historical prospectus, the entity, attribute and attribute value in the segmentation question should be generated according to the prospectus, so a question template can be generated first. As long as the entity, attribute and attribute value are filled in the question template, an inquiry letter for the prospectus can be generated. For example, the detailed question in the historical inquiry letter is "In 2018, 2019 and 2020, the company's overseas sales revenue accounted for 43.49%, 44.71% and 23.57% of the main business income respectively. Please explain the reason for the decline in main business income", then the question template after removing entities, attributes and attribute values and filling in placeholders is "In [year], [year] and [year], the company's overseas sales revenue accounted for [POOSITMBI], [POOSITMBI] and [POOSITMBI] respectively. Please explain the reason for the decline in main business income."
[0120] Afterwards, the triggering conditions of the audit points are determined based on the segmented questions, and a fourth directory is obtained that represents the correspondence between the audit points, inquiry directions, question templates and triggering conditions. The triggering conditions are used to determine whether to inquire about the audit points.
[0121] Specifically, the triggering conditions of the audit points are divided into two types: rule mode and content mode. The triggering conditions of the rule mode include:
[0122] An attribute value threshold is preset based on the attribute value of the detailed question. If the attribute value in the historical prospectus does not meet the attribute value threshold, the audit points are determined to be triggered. The attribute value threshold can be obtained through formulas, regular expressions and four arithmetic operations. The attribute values in the detailed questions in the historical inquiry letters correspond to the attribute values of the corresponding audit points in the historical prospectuses.
[0123] The trigger conditions for content mode include:
[0124] By using historical inquiry letters and historical prospectuses as training samples for machine learning training, a binary classification model is obtained to determine whether there are abnormal audit points. When the target prospectus is input into the binary classification model and the conclusion is that there are abnormal audit points, the audit points are determined to be triggered.
[0125] Machine learning is performed on the inquiry background paragraph and historical prospectuses to obtain summary rewriting rules for generating the inquiry background paragraph.
[0126] Specifically, when the content in the prospectus meets the triggering conditions of the audit points, an inquiry background paragraph needs to be generated in the inquiry letter. Therefore, the inquiry background paragraph and the historical prospectus are used as training samples for machine learning training to obtain the corresponding relationship between the inquiry background paragraph and the content in the historical prospectus;
[0127] A summary rewriting rule for generating the inquiry background paragraph is determined based on the corresponding relationship. The summary rewriting rule is used to rewrite the summary of the text paragraphs in the historical prospectus that have a corresponding relationship with the inquiry background paragraph to generate the inquiry background paragraph.
[0128] For example, when an audit point is triggered, first identify the paragraph in the prospectus that triggers the audit point, use the inquiry direction corresponding to the audit point as the keyword of the inquiry background paragraph, and preferentially extract sentences containing keywords from the prospectus. Identify the pronouns in these sentences, replace the anaphora with the antecedent to make the sentence more fluent, thereby generating an inquiry background paragraph.
[0129] The rule dictionary is determined to include the fourth directory and summary rewriting rules, and a target inquiry letter corresponding to the target prospectus is generated according to the rule dictionary.
[0130] 203. Obtain the text vector of the target prospectus.
[0131] Since the input of the target generation model is a text vector, it is necessary to first use a text algorithm to obtain the text vector of the target prospectus so that the target generation model can generate a target inquiry letter corresponding to the target prospectus based on the text vector of the target prospectus.
[0132] 204. Input the text vector of the target prospectus into the target generation model, and the target generation model outputs the target inquiry letter corresponding to the target prospectus according to the rule dictionary.
[0133] After the text vector of the target prospectus is input into the target generation model, the rule dictionary is matched with the text vector of the target prospectus. When the triggering conditions of the audit points are met, the rule dictionary performs similarity matching and keyword matching based on the text vector for the paragraph in the target prospectus that triggers the audit points. If the paragraph meets a certain question template, the question template is filled with entities, attributes and attribute values to generate segmented questions. At least one segmented question constitutes the inquiry question paragraph of the target inquiry letter. The inquiry background paragraph is generated according to the summary rewriting rules. Specifically, the inquiry direction corresponding to the audit points is used as the keyword of the inquiry background paragraph. Sentences containing keywords are preferentially extracted from the prospectus, the pronouns in these sentences are identified, and the anaphora is replaced with the antecedent to make the sentence more fluent, thereby generating the inquiry background paragraph. Among them, the audit points, inquiry question paragraphs and inquiry background paragraphs have a corresponding relationship. A target inquiry letter includes at least one set of audit points, inquiry question paragraphs and inquiry background paragraphs.
[0134] In an embodiment of the present application, a target generation model for generating a target inquiry letter is obtained by performing machine learning on historical inquiry letters and historical prospectuses. The target generation model stores a rule dictionary for generating a target inquiry letter, so that when the target prospectus is input into the target generation model, the target generation model can output a target inquiry letter, thereby reducing the manpower and time costs required for pure manual writing.
[0135] The following is a description of an inquiry letter generation system provided in an embodiment of the present application. Figure 3 An implementation of an inquiry letter generation system provided in an embodiment of the present application includes:
[0136] An acquisition unit 301 is used to acquire a target generation model, the target generation model is obtained by training an initial generation model through machine learning using historical inquiry letters and historical prospectuses, the target generation model stores a rule dictionary for generating a target inquiry letter, the historical inquiry letters and historical prospectuses have a corresponding relationship, and the historical inquiry letters are generated based on the historical prospectuses, and the rule dictionary is obtained by performing machine learning on the historical inquiry letters and historical prospectuses and combining manual experience;
[0137] The acquisition unit 301 is further used to acquire the text vector of the target prospectus;
[0138] The input unit 302 is used to input the text vector of the target prospectus into the target generation model, and the target generation model outputs the target inquiry letter corresponding to the target prospectus according to the rule dictionary.
[0139] The functions and processes performed by each unit in the inquiry letter generation system of this embodiment are similar to those of the aforementioned Figure 1 to Figure 2 The functions and processes performed by the inquiry letter generation system are similar and will not be repeated here.
[0140] Figure 4 It is a structural diagram of the inquiry letter generating device provided in the present application. The inquiry letter generating device 400 may include one or more central processing units (CPU) 401 and a memory 405. The memory 405 stores one or more application programs or data.
[0141] The memory 405 may be a volatile storage or a persistent storage. The program stored in the memory 405 may include one or more modules, each of which may include a series of instruction operations in the inquiry letter generating device. Furthermore, the central processor 401 may be configured to communicate with the memory 405, and execute a series of instruction operations in the memory 405 on the inquiry letter generating device 400.
[0142] The inquiry letter generating device 400 may also include one or more power supplies 402, one or more wired or wireless network interfaces 403, one or more input and output interfaces 404, and / or one or more operating systems, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.
[0143] The CPU 401 can execute the aforementioned Figure 1 to Figure 2 The operations performed by the inquiry letter generation system in the illustrated embodiment will not be described in detail here.
[0144] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0145] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.
[0146] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0147] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0148] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk and other media that can store program code.
Claims
1. A method for generating an inquiry letter, It is characterized in that include: Use a classification algorithm to decompose the historical inquiry letter into inquiry background paragraphs and inquiry question paragraphs, and identify the subdivided questions contained in the inquiry question paragraphs, wherein one audit point corresponds to at least one inquiry background paragraph and at least one inquiry question paragraph, and the inquiry background paragraph and the inquiry question paragraph corresponding to the same audit point have an associated relationship; Determine the audit points corresponding to the subdivided questions, wherein one audit point corresponds to at least one subdivided question, and form a first directory indicating the correspondence between the audit points and the subdivided questions; Obtaining keywords contained in the segmented questions, and using the keywords as the inquiry direction of the segmented questions; The subdivided questions corresponding to each of the audit points are clustered according to the inquiry direction to obtain a second directory representing the correspondence between the audit points, the inquiry directions and the subdivided questions, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one subdivided question; Input the subdivided questions into a neural network algorithm model to generate question templates, and obtain a third directory representing the correspondence between the audit points, the inquiry directions, and the question templates, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one question template; Determine the triggering condition of the audit point according to the subdivided question, and obtain a fourth directory representing the corresponding relationship between the audit point, the inquiry direction, the question template and the triggering condition, wherein the triggering condition is used to determine whether to inquire about the audit point; Performing machine learning on the inquiry background paragraph and historical prospectuses to obtain summary rewriting rules for generating the inquiry background paragraph; The rule dictionary includes the fourth directory and the summary rewriting rule; Obtaining a target generation model, wherein the target generation model is obtained by training an initial generation model through machine learning using the historical inquiry letters and the historical prospectus, wherein the target generation model stores the rule dictionary for generating the target inquiry letter, the historical inquiry letter and the historical prospectus have a corresponding relationship, and the historical inquiry letter is generated according to the historical prospectus; Get the text vector of the target prospectus; Inputting the text vector of the target prospectus into the target generation model, and the target generation model outputting the target inquiry letter corresponding to the target prospectus according to the rule dictionary; The inputting the text vector of the target prospectus into the target generation model comprises: after inputting the text vector of the target prospectus into the target generation model, matching the rule dictionary with the text vector of the target prospectus, when any triggering condition of the audit points is met, the rule dictionary performs text vector-based similarity matching and keyword matching on the paragraph in the target prospectus that triggers the audit points, if the paragraph matches any of the question templates, filling entities, attributes and attribute values in the question template to obtain subdivided questions constituting the inquiry question paragraph of the target inquiry letter, and determining the inquiry question paragraph of the target inquiry letter based on the subdivided questions constituting the inquiry question paragraph of the target inquiry letter; When the triggering conditions of any of the audit points are met, the inquiry direction corresponding to the audit point is obtained, and the sentence containing the inquiry direction corresponding to the audit point is extracted from the target prospectus, and the inquiry direction corresponding to the audit point in the sentence is replaced with the corresponding antecedent, so as to obtain the inquiry background paragraph of the target inquiry letter; A target inquiry letter based on any of the audit points, the inquiry question paragraph of the target inquiry letter, the inquiry background paragraph of the target inquiry letter, and the correspondence between the three constitutes a target inquiry letter.
2. The method for generating an inquiry letter according to claim 1, It is characterized in that Before acquiring the target generation model, the method further includes: Obtaining the historical inquiry letter and the historical prospectus, and using a text vector algorithm to obtain text vectors of the historical inquiry letter and the historical prospectus; The text vectors of the historical inquiry letters and the historical prospectuses are input into the initial generation model as training samples, and the initial generation model is subjected to machine learning training using the training samples to obtain the target generation model, in which the rule dictionary is stored.
3. The method for generating an inquiry letter according to claim 1, It is characterized in that The obtaining of the keywords contained in the segmented question and using the keywords as the inquiry direction of the segmented question includes: Performing word segmentation matching based on word granularity on the segmented problem to obtain a preprocessed word segmentation set, and using a text vector algorithm and a text classification algorithm to process the preprocessed word segmentation set to obtain a keyword seed word library; Expanding the keyword seed word library with a synonym dictionary based on artificial experience to obtain a keyword word library; or using a text vector algorithm to calculate the similarity of words in the keyword seed word library to expand the keyword seed word library to obtain the keyword word library; Keywords included in the segmented question are obtained according to the keyword thesaurus, and the keywords are used as the query direction of the segmented question.
4. The method for generating an inquiry letter according to claim 1, It is characterized in that The step of inputting the segmentation problem into the neural network algorithm model to generate a problem template comprises: Using a text vector algorithm to obtain a text vector for the segmented problem, inputting the text vector for the segmented problem as a training sample into an initial text recognition model for machine learning training, and obtaining a target text recognition model, wherein the target text recognition model stores a recognition method for entity text, attribute text, and attribute value text in the segmented problem; The entity text, the attribute text and the attribute value text in the segmented question are replaced with placeholders to obtain the question template.
5. The method for generating an inquiry letter according to claim 4, It is characterized in that The triggering conditions for determining the audit key points according to the segmented issues include: The triggering conditions of the audit points are divided into two types: rule mode and content mode. The triggering conditions of the rule mode include: Presetting the attribute value threshold according to the attribute value of the detailed question, and determining to trigger the audit key point if the attribute value in the historical prospectus does not meet the attribute value threshold, wherein the attribute value in the detailed question in the historical inquiry letter corresponds to the attribute value of the corresponding audit key point in the historical prospectus; The triggering conditions of the content mode include: A binary classification model for determining whether there are abnormal audit points is obtained by performing machine learning training using the historical inquiry letters and the historical prospectuses as training samples; When the target prospectus is input into the binary classification model and a conclusion is drawn that abnormal audit points exist, it is determined that the audit points are triggered.
6. The method for generating an inquiry letter according to claim 1, It is characterized in that The method of performing machine learning on the inquiry background paragraph and the historical prospectus to obtain a summary rewriting rule for generating the inquiry background paragraph includes: Perform machine learning training using the inquiry background paragraph and the historical prospectus as training samples to obtain a correspondence between the inquiry background paragraph and the content in the historical prospectus; A summary rewriting rule for generating the inquiry background paragraph is determined according to the corresponding relationship, and the summary rewriting rule is used to summarize the text paragraphs in the historical prospectus that have a corresponding relationship with the inquiry background paragraph to generate the inquiry background paragraph.
7. A system for generating inquiry letters, It is characterized in that include: an acquisition unit, configured to acquire a target generation model, wherein the target generation model is obtained by performing machine learning training on an initial generation model using historical inquiry letters and historical prospectuses, wherein a rule dictionary for generating a target inquiry letter is stored in the target generation model, the historical inquiry letters have a corresponding relationship with the historical prospectuses, and the historical inquiry letters are generated according to the historical prospectuses; The acquisition unit is further used to acquire the text vector of the target prospectus; An input unit, configured to input the text vector of the target prospectus into the target generation model, and the target generation model outputs a target inquiry letter corresponding to the target prospectus according to the rule dictionary; The acquisition unit is specifically used to use a classification algorithm to decompose the historical inquiry letter into an inquiry background paragraph and an inquiry question paragraph, and identify the subdivided questions contained in the inquiry question paragraph, wherein one audit point corresponds to at least one inquiry background paragraph and at least one inquiry question paragraph, and the inquiry background paragraph and the inquiry question paragraph corresponding to the same audit point have an associated relationship; Confirming the audit points corresponding to the subdivided questions, wherein one audit point corresponds to at least one subdivided question, and forming a first directory indicating the correspondence between the audit points and the subdivided questions; Obtaining keywords contained in the segmented questions, and using the keywords as the inquiry direction of the segmented questions; The subdivided questions corresponding to each of the audit points are clustered according to the inquiry direction to obtain a second directory representing the correspondence between the audit points, the inquiry directions and the subdivided questions, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one subdivided question; Input the subdivided questions into a neural network algorithm model to generate question templates, and obtain a third directory representing the correspondence between the audit points, the inquiry directions, and the question templates, wherein one audit point corresponds to at least one inquiry direction, and one inquiry direction corresponds to at least one question template; Determine the triggering condition of the audit point according to the subdivided question, and obtain a fourth directory representing the corresponding relationship between the audit point, the inquiry direction, the question template and the triggering condition, wherein the triggering condition is used to determine whether to inquire about the audit point; Performing machine learning on the inquiry background paragraph and the historical prospectus to obtain a summary rewriting rule for generating the inquiry background paragraph; The rule dictionary includes the fourth directory and the summary rewriting rule; The input unit is specifically used to input the text vector of the target prospectus into the target generation model, and then match the rule dictionary with the text vector of the target prospectus. When the triggering condition of any of the audit points is met, the rule dictionary performs text vector-based similarity matching and keyword matching on the paragraph in the target prospectus that triggers the audit point. If the paragraph matches any of the question templates, then fill in the entity, attribute and attribute value in the question template to obtain the subdivided questions constituting the inquiry question paragraph of the target inquiry letter, and determine the inquiry question paragraph of the target inquiry letter based on the subdivided questions constituting the inquiry question paragraph of the target inquiry letter; When the triggering conditions of any of the audit points are met, the inquiry direction corresponding to the audit point is obtained, and the sentence containing the inquiry direction corresponding to the audit point is extracted from the target prospectus, and the inquiry direction corresponding to the audit point in the sentence is replaced with the corresponding antecedent, so as to obtain the inquiry background paragraph of the target inquiry letter; A target inquiry letter based on any of the audit points, the inquiry question paragraph of the target inquiry letter, the inquiry background paragraph of the target inquiry letter, and the correspondence between the three constitutes a target inquiry letter.
8. A device for generating an inquiry letter, It is characterized in that include: CPU, memory and input / output interface; The memory is a short-term storage memory or a persistent storage memory; The central processing unit is configured to communicate with the memory and execute instructions in the memory to perform the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, It is characterized in that The method comprises instructions, which, when executed on a computer, cause the computer to execute the method according to any one of claims 1 to 6.
Citation Information
Patent Citations
Question data generation method and apparatus, computer apparatus, and storage medium
CN109543007A
Course trial query generation method and device for generating Seq2Seq model based on pointer, and medium
CN112417155A
Natural language question answering method and apparatus, and server
WO2018149326A1
