A method and device for automatically generating audit suggestions in an audit report

Abstract

The present invention discloses a method for automatically generating audit suggestions in an audit report. First, the audit suggestion statements in previous audit reports are collected to construct an audit suggestion library; then, the audit problems pointed out in the audit report are encoded to obtain their embedding vectors in the problem space; then, a mapper is used to map the encoding of the audit problems in the problem space to the suggestion space; finally, the semantic model and the mapper are jointly trained to learn the parameters therein. When writing a new audit report, the problems found in the audit are input into the semantic model and the mapper to obtain the embedding vector of the problem in the suggestion space, and several suggestion statements closest to the embedding vector are selected from the audit suggestion library to generate the audit suggestions in the report. The present invention can make full use of the rigorous and standardized audit suggestion statements in the existing audit reports accumulated over a long period of time, and automatically generate accurate and professional audit suggestions based on the problems found in the audit.

Description

Technical Field

[0001] The present invention relates to a method and device for automatically generating audit suggestions in an audit report, and belongs to the fields of information retrieval, natural language processing, and auditing. Technical Background

[0002] As the last step of the entire audit process, the audit report conveys the auditor's findings and is crucial to the audit business. In recent years, artificial intelligence, especially natural language processing technology, has been increasingly widely used in the auxiliary writing and automatic generation of audit reports to improve the standardization and rigor of audit reports and reduce the tediousness of report writing.

[0003] In the process of writing audit reports, auditors need to make targeted and actionable suggestions based on the problems found in the audit, and text generation technology is a method to generate audit suggestions. Many text generation methods based on deep learning have been used to automatically generate poetry, prose, etc., but they cannot be applied to automatically generate rigorous audit suggestions in audit reports. The reason is that audit suggestions must be based on objective facts and provided according to the problems found in the audit, and cannot be written freely like poetry or prose. In recent years, large language models (LLMs) such as GPT-4, Llama2, and Wenxinyiyan have shown excellent understanding and creativity, and can generate coherent texts that conform to human reading habits based on prompts, but there are also hallucination problems, and they cannot be directly used in the highly professional and serious scenario of audit suggestion generation. Therefore, due to the high requirements of audit reports on authenticity, accuracy, standardization, and operability, there are still many gaps in the application of text generation technology in audit suggestion writing. Summary of the invention

[0004] To solve the above problems, the present invention proposes a method and device for automatically generating audit recommendations based on text recommendations. Since the characterization of audit issues and the issuance of recommendations are based on laws, regulations and audit standards, the scope of the recommended content is relatively limited, so that a large number of ready-made audit recommendations accumulated in previous audit work can be reused in new audit reports. Therefore, the present invention will use existing audit reports to build an audit recommendation library containing recommendation statements, and then learn the semantic information of audit issues in the audit reports, and based on this, select appropriate recommendation statements to automatically generate audit recommendations in the audit reports. The technical problem to be solved by the present invention is how to accurately extract semantic information from audit issues, make full use of the semi-structured characteristics of audit issues to deepen the semantic model's understanding of audit issues, so as to generate rigorous and targeted audit recommendations. First, collect the audit suggestion statements in the previous audit reports, encode them using models such as Bert, and build an audit suggestion library; then, use the semantic model to encode the audit problems pointed out in the audit report to obtain their embedded representation in the problem space; then, use the mapper to map the encoding of the audit problems in the problem space to the suggestion space to facilitate the calculation of the distance with the encoded suggestion statements in the audit suggestion library; finally, use the existing audit reports to build a training set, and jointly train the above semantic model and mapper to learn the parameters. When writing a new audit report, input the problems found in the audit into the semantic model and mapper to obtain the embedded representation of the problem in the suggestion space, select several suggestion statements closest to the embedded representation from the audit suggestion library, and generate the audit suggestions in the report.

[0005] The present invention first discloses a method and device for automatically generating audit recommendations in an audit report, which learns semantic information from various parts of a semi-structured audit question, selects appropriate recommendation statements from an audit recommendation library according to the question semantics, and generates audit recommendations in the audit report. The main steps of the method of the present invention include:

[0006] Step S1, establish a training set, which consists of audit questions and their corresponding audit suggestions in the audit report, segment the audit questions and audit suggestions and use the Word2Vec model to encode the words. For the sentences in the audit suggestions, further input the word encoding contained in them into the Bert model to obtain the encoding vector of the suggestion sentence;

[0007] Step S2, using the question semantics learning model to encode the audit question, specifically including sentence encoding, question title encoding, question description encoding, question qualitative basis encoding, and encoding integration steps, to obtain its embedding vector in the question space. The specific steps are:

[0008] Step SS1, extract the main information words from the audit question and generate the main information word set I = {w1, w2, ..., w m}, where m is the number of information words, whose sources include but are not limited to: words in the question title, audit question topic description words obtained by using the topic model, and using methods such as TF-IDF or Textrank to sort the words in the audit question and extract the top words, etc.;

[0009] Step SS2: Input the words that make up the sentence into the LSTM network for sentence encoding. Suppose the xth sentence W x The word set is {w x,1 ,w x,2 ,…,w x,n}, n is the number of words in the sentence, and the calculation process of inputting the words into the LSTM module for sentence encoding is as follows:

[0010]

[0011] Among them, h t and h t-1 are the hidden vectors output by the LSTM module at steps t and t-1, respectively. Represents the word w that is input into the LSTM module at step t x,t The embedding vector obtained by word encoding, θ represents the parameter set to be learned in the LSTM module;

[0012] Step SS3, using the main information word to calculate the weight of each word in the sentence in the sentence encoding, first, calculate the similarity between each word and the main information word, the method is as follows:

[0013]

[0014] in, Indicates the main information word w i The embedding vector obtained by word encoding, Represents the yth word w in the xth sentence x,y The embedding vector obtained by word encoding, T represents transposition. Then, the softmax function is used to assign weights to each word as follows:

[0015]

[0016] in, Represents the yth word w y The obtained weights, EXP(·), represent the natural exponential function.

[0017] Step SS4, based on the weight of each word in the sentence, combined with the output of each step of the LSTM network, calculate the encoding of the sentence, the method is as follows:

[0018]

[0019] Among them, Δ s represents the sentence vector obtained after encoding sentence s, h y Indicates that the LTSM unit receives the word vector The hidden vector output after input.

[0020] In step SS5, the embedding vectors of the words constituting the question title are sequentially input into the LSTM unit, and the final output is used as the embedding vector of the question title encoding, as follows:

[0021]

[0022]

[0023] in, is the tth word in the question title, i.e., the word w that is input into the LSTM unit at step t t The embedding vector obtained by word encoding, is the latent vector output by the corresponding LSTM unit, θ represents the parameter set to be learned in the LSTM module, and shares parameters with the LSTM unit in step S22, r is the number of words in the title, i.e., w r is the last word of the title, Δ T The final encoded embedding vector for the question title.

[0024] Step SS6, assuming that the embedding vectors corresponding to the sentences constituting the problem description D are (obtained from steps SS1-SS4), l is the number of sentences in D, and the sentence embedding vectors are input into the gated recurrent unit in sequence:

[0025]

[0026] Among them, t and t-1 Respectively represent the hidden vectors of the GRU unit at step t and step t-1, represents the sentence vector input to the GRU unit at step t, is the parameter set of the GRU unit.

[0027] Step SS7, using the question title as the query vector, calculates the weight of each sentence in the question description as follows:

[0028]

[0029] in, represents the weight obtained by the ith statement, T represents transpose, and EXP(·) represents the natural exponential function.

[0030] Step SS8: Based on the weights of the sentences in the problem description and the output of each step of the GRU unit, the embedding vector Δ of the problem description is calculated. D , the method is as follows:

[0031]

[0032] Among them, i For input statement , which corresponds to the hidden vector output by the GRU unit.

[0033] Step SS9, encode the common sentences in the qualitative basis in the same way as the sentence encoding method in the problem description, and obtain the embedding vector Δ of the common sentence part of the problem qualitative basis J′ ;

[0034] Step SS10, for the names of the violated laws and regulations mentioned in the qualitative basis of the problem, calculate the corresponding embedding vector in advance through pre-coding. The specific method is: extract the names of laws and regulations and their corresponding audit questions and suggestions from the published audit cases, use the Word2Vec model to encode the words appearing in the audit questions and suggestions, and then calculate the mean of the encoded word embedding vector to obtain the corresponding legal and regulatory name encoding vector Δ R .

[0035] Step SS11, encoding the qualitative basis of the audit question to obtain the embedding vector Δ J , the specific method is:

[0036] Δ J =AVG(Δ J′ ,Δ R )

[0037] Among them, AVG means averaging the vector.

[0038] Step SS12: Integrate the problem title, problem description, and problem qualitative basis to form an embedding vector Δ of the complete audit problem encoding in the problem space P , the method is as follows:

[0039]

[0040] in, Represents the concatenation of vectors.

[0041] Step S3: Use a mapper to map the embedding vector of the audit question in the problem space to the audit suggestion space. The mapper is composed of a multi-layer perceptron (MLP), which includes multiple hidden layers and a linear activation layer. The specific composition and mapping method are as follows:

[0042] Δ PS =LIN(WL …ReLU(W2ReLU(W1Δ P +b1)+b2)+…+b L )

[0043] Among them, Δ PS Represents the audit question embedding vector Δ in the problem space P The embedding vector after mapping to the audit suggestion space, L represents the number of layers of the multilayer perceptron, W * and b * They represent the weight matrix and bias vector respectively, ReLU(·) represents the linear rectification function, and LIN(·) represents the linear activation function.

[0044] Step S4, establish an optimization objective function, minimize the sum of the distances between the embedding vectors of all audit questions encoded in the audit suggestion space and the embedding vectors of their corresponding audit suggestions, and train the parameters in the problem semantic learning model and the mapper. The optimization objective function is designed to minimize the distance between the audit question embedding vector in the audit suggestion space and the corresponding audit suggestion vector. The audit questions and their corresponding audit suggestion samples given in the training set are trained to learn the parameter values ​​in the semantic learning module and the mapper. The specific method is to establish the following loss function:

[0045]

[0046] in, Represents the training set, which contains M samples, each of which is an audit question P in an audit report i and its corresponding audit recommendation statement S i , For audit question P i The embedding vector in the audit suggestion space obtained after the semantic learning module and mapper conversion, Suggested statement for audit i The embedding vector obtained after encoding by the Bert model, Dist(·) represents the distance function, Θ represents the parameter set involved in the semantic learning and mapping process, λ represents the normalization parameter, and ||.||2 represents the L2 norm. The parameters in the parameter set Θ are trained using the training set constructed by the existing report to minimize the loss function. The learned problem semantic learning model and mapper will be used for subsequent new audit problem encoding.

[0047] Step S5, collecting audit suggestions in existing audit reports, encoding the statements in the audit suggestions using models such as Bert, and establishing an audit suggestion library;

[0048] Step S6, when writing a new audit report, use the trained problem semantic learning model and mapper to encode the audit problem, obtain the representation vector of the audit problem in the audit suggestion space, select several vectors closest to the vector from the audit suggestion library, and form the audit suggestion from the corresponding suggestion statements.

[0049] The present invention also discloses a device for automatically generating audit suggestions from an audit report, comprising an audit report parsing unit, an audit question semantic learning unit, an audit question mapping unit, an audit suggestion library unit, and an audit suggestion selection unit. Among them: The audit report parsing unit analyzes the audit report, extracts audit questions from it, and further parses to obtain question titles, questions, and suggestions.

[0050] The Word2Vec model is used to perform an initial analysis of the words and sentences in the topic description, qualitative basis, and violated laws and regulations.

[0051] Step coding;

[0052] The audit question semantics learning unit receives the result of the audit report parsing unit and learns the audit question semantics according to step S2

[0053] and coding the audit questions;

[0054] The audit question mapping unit maps the encoding vector output by the audit question semantic learning unit to the audit construction vector according to step S3.

[0055] meeting space;

[0056] The audit suggestion library unit stores audit suggestion statements extracted from existing audit reports. The statements are encoded using the Bert model.

[0057] Stored in the form of <statement, statement vector> pairs, which facilitates the calculation of the distance between vectors;

[0058] The audit suggestion selection unit selects the audit suggestion from the audit suggestion library according to the encoding vector of the audit question in the suggestion space output by the mapper.

[0059] A number of vectors closest to the vector are selected, and the corresponding suggestion statements are automatically combined to generate the audit suggestion part in the audit report.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows: it can make full use of the semi-structured characteristics of audit questions, consider the importance of different parts of audit questions in question semantic learning, so as to understand audit questions more accurately and select more targeted audit recommendation statements. The present invention generates audit recommendations in a recommendation manner, effectively avoiding model hallucinations and the randomness of generated text, and can generate rigorous and usable audit recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is an overall flow chart of the method of the present invention.

[0062] Figure 2 It is a structural schematic diagram of the device of the present invention.

[0063] Figure 3 Schematic diagram of audit problems in an embodiment of the present invention.

[0064] Figure 4 Schematic diagram of audit suggestion in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is described in detail below with reference to the accompanying drawings:

[0066] Figure 2 The device of the present invention includes an audit report parsing unit 101, an audit question semantic learning unit 102, an audit question mapping unit 103, an audit suggestion library unit 104, and an audit suggestion selection unit 105. Among them:

[0067] The audit report parsing unit analyzes the audit report, extracts audit questions from it, and further parses the question title, question

[0068] The Word2Vec model is used to perform an initial analysis of the words and sentences in the topic description, qualitative basis, and violated laws and regulations.

[0069] Step coding;

[0070] The audit question semantics learning unit receives the result of the audit report parsing unit and learns the audit question semantics according to step S2

[0071] and coding the audit questions;

[0072] The audit question mapping unit maps the encoding vector output by the audit question semantic learning unit to the audit construction vector according to step S3.

[0073] meeting space;

[0074] The audit suggestion library unit stores audit suggestion statements extracted from existing audit reports. The statements are encoded using the Bert model.

[0075] Stored in the form of <statement, statement vector> pairs, which facilitates the calculation of the distance between vectors;

[0076] The audit suggestion selection unit selects the audit suggestion from the audit suggestion library according to the encoding vector of the audit question in the suggestion space output by the mapper.

[0077] A number of vectors closest to the vector are selected, and the corresponding suggestion statements are automatically combined to generate the audit suggestion part in the audit report.

[0078] Figure 1 The process of training the semantic learning module and mapper, and using them to select suggestion statements from the audit suggestion library to generate audit suggestions is shown. The specific steps are as follows:

[0079] Step S1, establish a training set, which consists of audit questions and their corresponding audit suggestions in the audit report, segment the audit questions and audit suggestions and use the Word2Vec model to encode the words. For the sentences in the audit suggestions, further input the word encoding contained in them into the Bert model to obtain the encoding vector of the suggestion sentence;

[0080] Step S2, using the question semantic learning model to encode the audit question. First, the sentences in the question are encoded based on the word encoding. Then, according to the structure of the audit question, the question title, question description, and the qualitative basis of the question are encoded respectively. Finally, the encoding of each part is integrated into a complete audit question encoding to obtain its embedding vector in the question space.

[0081] Step S3, using a mapper to map the embedding vector of the audit question in the question space to the audit suggestion space;

[0082] Step S4, establishing an optimization objective function to minimize the sum of the distances between the embedding vectors of all audit questions in the training set encoded in the audit suggestion space and the embedding vectors of their corresponding audit suggestions, and training the problem semantic learning model and parameters in the mapper;

[0083] Step S5, collecting audit suggestions in existing audit reports, encoding the statements in the audit suggestions using models such as Bert, and establishing an audit suggestion library;

[0084] Step S6, when writing a new audit report, use the trained semantic learning model and mapper to encode the audit question, obtain the representation vector of the audit question in the audit suggestion space, select several vectors closest to the vector from the audit suggestion library, and form the audit suggestion from the corresponding suggestion statements.

[0085] Example

[0086] The following is a detailed embodiment of the present invention. Figure 3 The audit questions shown generate audit recommendations:

[0087] The audit suggestion library contains 5 suggestion statements, namely:

[0088] 1. Prepare and submit budget estimates and final accounts scientifically and reasonably, strictly implement departmental budgets, and all expenditures must be listed according to the approved budget items and amounts to fully and comprehensively reflect the true situation of fiscal revenue and expenditure, and effectively maintain the seriousness and binding force of departmental budgets.

[0089] 2. Standardize the preparation of final accounts and ensure the authenticity of final accounts. Ensure that accounting information is true and complete and improve the quality of accounting information.

[0090] 3. Strengthen the management of fixed assets, establish and improve the fixed asset management system, and strictly handle the purchase, scrapping and transfer procedures of fixed assets in accordance with regulations.

[0091] 4. Process accounts in a timely manner, strengthen daily supervision, implement management responsibilities, institutionalize and standardize fixed asset management, and ensure that assets are authentic and complete.

[0092] 5. Improve the financial management system and internal control system, correctly account for accounting business, strictly implement payment approval procedures, standardize the content of original vouchers, ensure the authenticity, validity and completeness of accounting information, and promote the standardization of financial basic work.

[0093] The corresponding sentence vectors obtained by Bert encoding are Y1, Y2, Y3, Y4, and Y5 respectively.

[0094] First, encode each statement in the audit recommendation according to steps SS1-SS4 to obtain the corresponding embedding vector. Next:

[0095] (1) Encode the title of the audit question: "Irregular management of fixed assets". According to the method in step SS5, the words in the sentence are input into the LSTM unit one by one. The latent vector output by the last unit is used as the embedding vector Δ of the title. T ;

[0096] (2) The description of the audit problem: "Fixed assets are mixed with other units. The audit inventory of fixed assets found that the office equipment of the Municipal Tourism Bureau was mixed with the office equipment of the Municipal Annual Pass Office and the Municipal Quality Supervision Institute." Encode it. According to the method of steps SS6-SS8, each sentence in the problem description is input into the GRU unit, and Δ T Calculate the weight of each sentence as the query vector and obtain the embedding vector Δ of the question description D ;

[0097] (3) General sentences in the basis for the characterization of audit issues: "Administrative units should regularly check and inventory the state-owned assets they own and use to ensure that their financial status is clear and that the accounts, cards and actual situation are consistent, so as to prevent the loss of state-owned assets."

[0098] If the state-owned assets that the administrative unit intends to occupy or use are rented or loaned out, they must be reported to the financial department at the same level for review and approval in advance. Without approval, they shall not be rented or loaned out. "Encode it according to the method described in step SS9, similar to the encoding of the audit description statement, to obtain the embedding vector Δ of the common statement in the audit qualitative basis J′ ;

[0099] (4) Encode the name of the law and regulation in the qualitative basis of the audit problem: "Interim Measures for the Administration of State-owned Assets of Administrative Units", and obtain the embedding vector of the name of the law and regulation in the qualitative basis of the audit according to the method described in step SS10.

[0100] Δ R ;

[0101] (5) Encode the qualitative basis of the audit question and, according to the method described in step SS11, J′ and Δ R Take the mean and obtain the embedding vector Δ of the audit qualitative basis J ;

[0102] (6) According to the method described in step SS12, the audit question title, description, and qualitative basis are concatenated to obtain the embedding vector Δ of the audit question. P .

[0103] Then, according to step S3, the audit question is embedded into the vector Δ P Input the mapper and obtain its embedding vector Δ in the audit suggestion space PS .

[0104] Finally, calculate Δ PS The distance between the suggestion statement vectors Y1, Y2, Y3, Y4, and Y5 in the audit suggestion library. Assuming that the two vectors with the closest distance are taken, the results are Y3 and Y4. For this audit problem, the final generated audit suggestion is as follows Figure 4 shown.

[0105] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for automatically generating audit recommendations in an audit report, characterized in that The problem semantic information is learned based on the paragraph structure of the audit problem described in the audit report, and then appropriate suggestion sentences are recommended from the audit suggestion library based on the problem semantics to automatically generate the audit suggestion part in the audit report; the steps include: Step S1, establish a training set, which consists of audit questions and their corresponding audit suggestions in the audit report, segment the audit questions and audit suggestions and use the Word2Vec model to encode the words. For the sentences in the audit suggestions, input the word encoding contained in them into the Bert model to obtain the encoding vector of the suggestion sentence; Step S2, using the question semantic learning model to encode the audit question, first, encode the sentence in the audit question based on the word encoding; then, according to the structure of the audit question, encode the question title, question description, and question qualitative basis respectively; finally, integrate the encoding of each part into a complete audit question encoding to obtain its embedding vector in the question space; Step S3, using a mapper to map the embedding vector of the audit question in the question space to the audit suggestion space; Step S4, establishing an optimization objective function to minimize the sum of the distances between the embedding vectors of all audit questions in the training set encoded in the audit suggestion space and the embedding vectors of their corresponding audit suggestions, and training the problem semantic learning model and parameters in the mapper; Step S5, collecting audit suggestions in existing audit reports, encoding the sentences in the audit suggestions, and establishing an audit suggestion library; Step S6, when writing a new audit report, use the trained problem semantic learning model and mapper to encode the audit problem, obtain the representation vector of the audit problem in the audit suggestion space, select several vectors closest to the vector from the audit suggestion library, and form the audit suggestion from the corresponding suggestion statements.

2. The method according to claim 1, characterized in that In step S2, the encoding method for the statement in the audit question is: input the embedding vector of the words that make up the statement into the long short-term memory network LSTM, then use the main information words to calculate the weight of each word in the statement, and finally weight the embedding vector of the statement. The specific steps are as follows: Step SS1, extract the main information words from the audit question and generate the main information word set I = {w1, w2, ..., w m }, where m is the number of information words, whose sources include but are not limited to: words in the question title, audit question topic description words obtained by using the topic model, and using the TF-IDF or Textrank method to sort the words in the audit question and extract the top words; Step SS2: Input the words that make up the sentence into the LSTM network for sentence encoding and define the words that make up the xth sentence W x The word set is {w x,1 ,w x,2 ,…,w x,n }, n is the number of words in the sentence, and the calculation process of inputting the words into the LSTM module for sentence encoding is as follows: Among them, h t and h t-1 are the hidden vectors output by the LSTM module at steps t and t-1, respectively. Represents the word w that is input into the LSTM module at step t x,t The embedding vector obtained by word encoding, θ represents the parameter set to be learned in the LSTM module; Step SS3, using the main information word to calculate the weight of each word in the sentence in the sentence encoding, first, calculate the similarity between each word and the main information word, the method is as follows: in, Indicates the main information word w i The embedding vector obtained by word encoding, Represents the yth word w in the xth sentence x,y The embedding vector obtained by word encoding, T represents transposition; then, the softmax function is used to assign weights to each word as follows: in, Represents the yth word w y The obtained weights, EXP(·) represents the natural exponential function; Step SS4, based on the weights of each word in the obtained sentence in the encoding, combined with the output of each step of the LSTM network, calculate the encoding of the sentence, the method is as follows: Among them, Δ s represents the sentence vector obtained after encoding sentence s, h y Indicates that the LTSM unit receives the word vector The hidden vector output after input.

3. The method according to claim 2, characterized in that In step S2, the encoding method for the question title is: directly input the words in the title into the LSTM module in order, and use the output of the last step as the embedding vector obtained by encoding the title. The specific steps are: In step SS5, the embedding vectors of the words constituting the question title are sequentially input into the LSTM unit, and the final output is used as the embedding vector of the question title encoding, as follows: in, is the tth word in the question title, i.e., the word w that is input into the LSTM unit at step t t The embedding vector obtained by word encoding, is the latent vector output by the corresponding LSTM unit, θ represents the parameter set to be learned in the LSTM module, which shares parameters with the LSTM unit in step SS2, and r is the number of words in the title, i.e., w r is the last word of the title, Δ T The final encoded embedding vector for the question title.

4. The method according to claim 3, characterized in that In step S2, the encoding method for the problem description is: the problem description sentence is re-encoded using the gated recurrent unit GRU encoding, and then the weight of each sentence in the problem description is calculated using the problem title, and then the embedding vector of the problem description is obtained by weighting; the embedding vectors corresponding to the sentences constituting the problem description D are defined as l is the number of statements in D. The specific steps are: Step SS6, input the sentence embedding vector into the gated recurrent unit in sequence: Among them, t and t-1 Respectively represent the hidden vector output by the GRU unit at step t and step t-1, represents the sentence vector input to the GRU unit at step t, is the parameter set of the GRU unit; Step SS7, using the question title as the query vector, calculates the weight of each sentence in the question description as follows: in, represents the weight obtained by the ith statement, T represents transposition, and EXP(·) represents the natural exponential function; Step SS8: Based on the weights of each sentence in the problem description and the output of each step of the GRU unit, the embedding vector Δ of the problem description is calculated. D , the method is as follows: Among them, i For input statement , which corresponds to the hidden vector output by the GRU unit.

5. The method according to claim 4, characterized in that In step S2, the coding of the basis for characterizing the problem includes the following steps: Step SS9, encode the common sentences in the qualitative basis in the same way as the sentence encoding method in the problem description, and obtain the embedding vector Δ of the common sentence part of the problem qualitative basis J′ ; Step SS10, for the names of the violated laws and regulations mentioned in the qualitative basis of the problem, calculate the corresponding embedding vector in advance through pre-coding. The specific method is: extract the names of laws and regulations and their corresponding audit questions and suggestions from the published audit cases, use the Word2Vec model to encode the words appearing in the audit questions and suggestions, and then calculate the mean of the encoded word embedding vector to obtain the corresponding legal and regulatory name encoding vector Δ R ; Step SS11, encode the qualitative basis of the audit question to obtain the embedding vector Δ J , the specific method is: D J =AVG(Δ J′ ,D R ) Among them, AVG means averaging the vector.

6. The method according to claim 5, characterized in that In step S2, the embedding vector of the audit question is formed by concatenating the embedding vectors obtained by encoding the various parts of the audit question. The steps are as follows: Step SS12: Integrate the problem title, problem description, and problem qualitative basis to form an embedding vector Δ of the complete audit problem encoding in the problem space P , the method is as follows: in, Represents the concatenation of vectors.

7. The method according to claim 1, characterized in that In step S3, a mapper is used to transform the representation vector of the audit question in the problem space into the suggestion space. The mapper is composed of a multi-layer perceptron MLP, which includes multiple hidden layers and a linear activation layer. The specific composition and mapping method are as follows: Δ PS =LIN(W L …ReLU(W2ReLU(W1Δ P +b1)+b2)+…+b L ) Among them, Δ PS Represents the audit question embedding vector Δ in the problem space P The embedding vector after mapping to the audit suggestion space, L represents the number of layers of the multilayer perceptron, W * and b * They represent the weight matrix and bias vector respectively, ReLU(·) represents the linear rectification function, and LIN(·) represents the linear activation function.

8. The method according to claim 1, characterized in that In step S4, the optimization objective function is designed to minimize the distance between the audit question embedding vector and the corresponding audit suggestion vector in the audit suggestion space. The audit questions and their corresponding audit suggestion samples given in the training set are trained to learn the parameter values ​​in the semantic learning module and the mapper. The specific method is to establish the following loss function: in, Represents the training set, which contains M samples, each of which is an audit question P in an audit report i and its corresponding audit recommendation statement S i , For audit question P i The embedding vector in the audit suggestion space obtained after the semantic learning module and mapper conversion, Suggested statement for audit i The embedded vector obtained after encoding by the Bert model, Dist(·) represents the distance function, Θ represents the parameter set involved in the semantic learning and mapping process, λ represents the normalization parameter, and ||.||2 represents the L2 norm; the parameters in the parameter set Θ are trained using the training set constructed using the existing report to minimize the loss function. The learned problem semantic learning model and mapper will be used for subsequent new audit problem encoding.

9. A device for automatically generating audit suggestions in an audit report, implementing the method for automatically generating audit suggestions in an audit report as claimed in claim 1, characterized in that The device comprises an audit report parsing unit, an audit question semantic learning unit, an audit question mapping unit, an audit suggestion library unit, and an audit suggestion selection unit, wherein: The audit report parsing unit analyzes the audit report, extracts audit issues from it, and further parses it to obtain the issue title, issue description, qualitative basis, and violated laws and regulations, and calls the Word2Vec model to perform preliminary encoding on the words and sentences therein; The audit question semantics learning unit receives the results of the audit report parsing unit, learns the audit question semantics and encodes the audit question; The audit question mapping unit maps the encoding vector output by the audit question semantic learning unit to the audit suggestion space; The audit suggestion library unit stores the audit suggestion statements extracted from the existing audit reports. The statements are encoded using the Bert model and stored in the form of <statement, statement vector> pairs. The audit suggestion selection unit selects several vectors closest to the encoding vector of the audit question in the suggestion space output by the mapper from the audit suggestion library, and automatically combines the corresponding suggestion statements to generate the audit suggestion part in the audit report.

Images