Intelligent auditing techniques for enterprise finance

3.1 Intelligent analysis algorithm for accounting vouchers

As an important part of the implementation of financial intelligence in enterprises, a clustering algorithm combining word co-occurrence and SOM neural network is used to complete the clustering of voucher summaries, which in turn identifies suspicious points for review and provides assistance to auditors’ judgment. The algorithm first obtains co-occurrence resources from the text set and vectorizes the text and then clusters the text using the optimized SOM neural network. Finally, the results based on the clustering are evaluated [18,19]. If the text set is denoted by D = { d 1 , d 2 , … , d n } , and the number of texts in the text set is n , then, the n text is denoted by d n . The corresponding word set contains m words and the word set in n text is denoted by the K = { k n 1 , k n 2 , … , k n m } . Assuming that there are co-occurring lexical groups in the same window cell, the inter-word co-occurrence of two words is calculated as shown in equation (1).

where R ( ( k n i , k n j ) ∣ k n i ) and R ( ( k n i , k n j ) ∣ k n j ) refer to conditional co-occurrence, which are the proportion of co-occurring phrases ( k n i , k n j ) appearing in the same window when the vocabulary occurs. The number of occurrences of the vocabulary k n i , k n j and ( k n i , k n j ) co-occurring phrases are denoted by N ( k n i ) , N ( k n j ) , and N ( k n i , k n j ) , respectively. β and λ are the weighting coefficients of the co-occurrence rates of the vocabulary k n j and k n i , respectively. To remove the co-occurrence phrases with low association, the TexRank algorithm is used to construct the co-occurrence graph. The co-occurrence network graph is shown in Figure 2. The feature words are depicted by nodes in the graph. The co-occurrence rate between words is represented by the edges between nodes. The co-occurrence threshold is set, and the connected edges are the inter-word co-occurrence rate above a certain threshold.

Figure 2

Co-occurrence network diagram.

If i represents a node in the cell window, the correlation degree of the node can be calculated by equation (2).

where the weights of the joint edge of nodes i and j are represented by W ( i , j ) . The sets of nodes j and i pointing to other adjacent points are represented by O ( j ) and I ( i ) , respectively. The damping coefficients are portrayed by δ . The whole text is selected as the window unit, the keywords with larger weights is obtained by the Term frequency Inverse document frequency (TF-IDF) algorithm, the inter-word co-occurrence rate of the keywords, and the construction of the co-occurrence network graph are calculated according to TexRank algorithm. The text vector is used to represent the co-occurrence resources. The extraction process of co-occurring phrases is shown in Figure 3, which mainly includes three steps: text pre-processing, extraction of the first K keywords, and statistics of co-occurrence rate between words. In the text pre-processing stage, the original text resource is first cleaned with data to remove the worthless text information, then the vocabulary of the document set is sliced, and the vocabulary set is obtained based on the word separation tools. Finally, the deactivated words in the set are removed [20,21]. When extracting the first K keywords, TF-IDF algorithm is applied to calculate the weight of the words and remove the low frequency words. The co-occurrence network is constructed by counting the co-occurrence rate among the words. The co-occurrence phrases of keywords with high co-occurrence rate are selected to form the co-occurrence resource COW = { c 1 , c 2 , … , c l } .

Figure 3

Schematic diagram of co-occurring phrase extraction.

As a competitive learning unsupervised neural network, SOM neural network has the advantages of high objectivity, high accuracy, and few parameters in clustering time and effect. The classical SOM network includes an input layer and an output layer. The input vector and the number of neurons in the input layer have the same dimension. The neurons in the output layer have weight vectors. The SOM network clustering process includes three stages: competition, cooperation, and weight adjustment. Cooperation and weight adjustment refers to the simulation of the inhibitory function that biological cells always produce on the rest of the cells when they are subjected to the stimulation process. The adjustment weights include the winning nodes and the nodes in the domain. The algorithm is prone to dead nodes in more distant regions during training. It is optimized by two methods [22,23]. The first method is to improve the SOM network weight initialization method matrix. The first step is to obtain the average vector of all samples, and then set the initial value of the weight vector as the sum of the average vector and the random number. The second way is to improve the learning rate function in the weight adjustment stage, including the initial stage and the later stage of clustering. In the early stage, exponential function is applied to increase the updating range of weights. In the later stage, linear function is used to optimize weights.

3.2 Intelligent analysis and processing of audit reports

The intelligent analysis of the audit report supports the auditors to deal with the problems in a targeted way. A classification algorithm (WTL-BiLSTM) is proposed to achieve text classification, which involves four parts, namely, text preprocessing, vectorization, neural network layer, and comprehensive classification evaluation. Figure 4 displays the particular flowchart.

Figure 4

Flowchart of the WTL-BiLSTM algorithm.

Vectorization of text is completed through Word2vec, TF-IDF, and LDA. The semantic features of short texts, as well as the integration of textual thematic features and lexical importance features are obtained [24]. Then, the semantic features of the text are obtained from two aspects, respectively, by BiLSTM to solve the gradient disappearance and gradient explosion. Word2vec model sets the window size to 5. The dimension of the word vector is 100, with a quantity of 168,823. The prior parameters of the word distribution in the topic and the topic distribution in the document are set to 0.01 in the LDA topic model. The number of topics is 100, and the number of iterations is 1,000 [25]. The structure of the BiLSTM includes input gates, forgetting gates, and output gates. The structure is shown in Figure 5.

Figure 5

Schematic diagram of the BiLSTM structure.

The function of the forgetting gate f t is to determine what historical information is discarded. h t − 1 and x t refer to the output of the previous element and the input of the current element, respectively. δ refers to the Sigmoid function. w f is the weight matrix of the forgetting gate and b refers to the bias vector. The output value of the function is in the range [ 0 , 1 ] . When the output value of the forgetting gate is 1, the historical data are retained. Otherwise, the data are deleted. The calculation formula is shown in equation (3).

The input gate function determines whether new information is entered and involves two components. One is new content obtained by the Sigmoid function and the other is the alternative content obtained by the Tanh function [26]. The output gate determines whether the information is output or not, and also involves two parts. One is the output content of the information determined by the Sigmoid function. The other is the final output information h t obtained by multiplying the Tanh function and o t . w o is the weight matrix of the forgetting gate. The calculation formula is shown in equation (4).

To complete the BiLSTM, forward and backward propagations are required. The former is to calculate the output value of each module, and the latter is to calculate the error value of each module, which can transfer the error value of the current time to the upper layer in reverse [27]. The input to the BiLSTM is a character feature vector. It is stitched together in real time with forward and reverse outputs to obtain a bi-directional feature sequence, which can be expressed as h → t = [ h → t ; h ← t ] ∈ R m [28]. The hidden layer output sequence is ( h 1 , h 2 , h 3 , … , h t ) ∈ R n × m processed by the Tanh function. Assume that the number of labels in the annotation system is M and the m-dimension vector is transformed by the dropout function to obtain k-dimension. The obtained sentence-level text features are p, denoted by p = ( p 1 , p 2 , … , p n ) ∈ R n × k . Each vector in the matrix is the score of the j label in the i classification. After the SoftMax normalization, the output expression of the BiLSTM layer is shown equation (5).

where both w and b refer to the training parameters of the model. h ( w , b ) ( x ( i ) ) refers to the predicted value y of the i sample. Each vector takes a range of values [ 0 , 1 ] and the dimension sums to 1. Marks and dimensions in the annotation system correspond to each other. The dimension value increases with the increase in labeling probability. The algorithm performance evaluation metrics include precision, accuracy, and F1-Score. F1-Score is a metric for evaluating the recall and accuracy of a classifier, as shown in equation (6).

where R and P represent recall and precision, respectively, β represents the ratio of precision to recall. The F1-Score is a commonly used metric to evaluate the performance of classifiers [15,16].

4.1 Performance analysis of the intelligent analysis algorithm for accounting vouchers

To verify the effectiveness of the intelligent audit method proposed in the study, the improved SOM algorithm was tested and compared with TF-IDF SOM and COWORDS SOM algorithms. The experimental parameter settings are shown in Table 1.

The summary of accounting vouchers corresponding to the accounting subject is selected for auditing. The voucher abstracts include five types of vouchers: payroll payable, accounts receivable, prepayments, short-term investments, and notes receivable. There are 1,485 voucher summaries in total, as shown in Figure 6. The TF-IDF algorithm is used to obtain 1,523 keywords. The word co-occurrence algorithm is used to obtain the top 30 keyword groups in each type of document, totaling 300 keywords. The main co-occurring phrases in the category of payroll payable are month-rent, rent-rent, and year-purchase. The main co-occurring phrases in the category of accounts receivable are register-social, account-cleared, and company-returned. In prepayments category, they are loan-pre-purchase, unit-prepaid, and account-contract. In the category of short-term investments, the main co-occurring phrases are rental-rent, shop-rent, and year-purchase. In the category of notes receivable, the main co-occurring phrases are notes-unit, draft-received, and discount-advance. The five document data are divided into training, testing, and validation sets in a 3:1:1 ratio, with a total of 891, 297, and 435 documents, respectively.

Figure 6

Summary of voucher details for five accounts.

The output layer of the SOM network is set to be a 2 × 3 two-dimensional matrix. The initial neighborhood radius and learning rate are 2 and 1. The performance of the improved SOM algorithm (COWORDS-ISOM), COWORDS-SOM, and TF-IDF-SOM is compared experimentally. The results are shown in Figure 7. The accuracy, recall, and F1 values of the COWORDS-ISOM algorithm are better than the other two algorithms. The accuracy of COWORDS-ISOM, COWORDS-SOM, and TF-IDF-SOM are 85.12, 82.35, and 79.27, respectively. The recall rates are 83.28, 83.14, and 76.34% and the F1 values are 84.85, 82.14, and 77.37%, respectively.

Figure 7

Performance of SOM algorithm before and after optimization.

4.2 Results of intelligent analysis of audit reports

The training data for the Word2vec model is more than 1.5 million economic texts. The final word vectors are obtained. The experimental data used in the study was more than 1,800 audit questions obtained from audit reports. The ratio of training set, testing set, and validation set is 3:1:1. It is finally found that most documents are composed of 100–200 lexical groups. Figure 8 depicts the precise distribution. After the document is standardized, the number of words in each document is 100. If the number exceeds 100, the redundant part is truncated. If the number is insufficient, additional supplement is required.

Figure 8

Details of the number of words in the document.

The experiments are conducted with the help of keras 2.1.1 to build a neural network, including an input layer, a BiLSTM layer, a network layer, a dropout layer, and a fully connected layer. The parameter settings are shown in Table 2.

The performance of the WTL-BiLSTM model, the word vector-oriented bi-directional LSTM model (Word2vec Bi LSTM, W-BiLSTM), and the word vector-oriented LSTM model (Word2vec LSTM, W-LSTM) are compared. The results are presented in Figure 9. On the whole, the three models have high accuracy, recall, and F1 value, which completes short text classification well. However, the WTL-BiLSTM model has better performance. The accuracy, recall, and F1 values are superior to the other two models, with values of 87.43, 87.88, and 87.66%, respectively. The corresponding values of the W-BiLSTM model are 86.76, 85.55, and 85.63%. The corresponding values of the W-LSTM model are 86.03, 82.98, and 84.01%, respectively. Therefore, the WTL-Bi-LSTM model proposed in the study has the best intelligent analysis effect of audit reports.

Figure 9

Comparison of the performance of the three models.

The existing intelligent audit methods commonly used are K-means clustering analysis and Intelligent audit method of fast correlation-based filter (FCBF) algorithm. To compare the application performance of the optimization model, the commonly used K-means and FCBF audit analysis methods are compared with WTL-BiLSTM. The comparison results are shown in Figure 10. The accuracy of K-means method is 83.79%. The accuracy of FCBF model is 85.63%. The accuracy of WTL-BiLSTM model is 87.25%, which is significantly higher than the other existing intelligent audit analysis methods. This shows that the method has good data analysis effect and can provide reference for further optimization of existing audit analysis methods.

Figure 10

Comparison of accuracy of three audit methods.