Chinese financial evaluation unit extraction method based on inter-character relationship

Abstract

A method for extracting Chinese financial evaluation units based on inter-character relationships includes: encoding Chinese financial text sentences at the encoding layer to obtain character vectors; feeding the character vectors and the inter-character relationship matrix into a graph convolutional neural network model, enhancing the Chinese character relationships between evaluation elements on the character vectors, and encoding the dependency relationships; constructing a grid of character pair relationships in the evaluation text, and differentiating the words on the rows and columns of the grid using two unit convolution kernels; performing convolution operations using image convolution kernels of different sizes in the row and column directions; for elements in the grid, fusing their corresponding word codes on the rows and columns and the dependency relationship codes between the two words, determining the element's label by decoding, and completing the extraction of the evaluation unit. The present invention can solve the problems of insufficient utilization of Chinese character relationships between evaluation elements, failure to distinguish different evaluation element types when representing grid elements, and insufficient utilization of Chinese character relationships within evaluation elements.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for extracting Chinese financial evaluation units based on inter-character relationships. Background Art

[0002] In recent years, with the development of the internet and the financial industry, the amount of financial information available online has become increasingly rich. This content includes a large amount of sentimental comments, such as users' subjective opinions on specific products or the financial industry. With the development of the financial sector and the advancement of internet technology, more and more financial text has appeared on the internet.

[0003] Financial commentary encompasses not only the financial and other non-financial information disclosed by listed companies themselves, but also various external environmental information influencing the companies, as well as interpretations of the companies and the external environment by major institutions and experts. The knowledge contained in financial commentary can significantly help interpret the operating conditions and environment of listed companies, possessing extremely high strategic value. Consequently, the mining of financial commentary texts is gaining increasing attention.

[0004] Extracting aspect-opinion pairs, also known as evaluation unit extraction, is a key task in aspect-based sentiment analysis. An evaluation unit is generally represented as <evaluation object, evaluation word>. Evaluation unit extraction involves pairing the evaluation object and the evaluation word. Extracting both together is crucial because the sentiment polarity of an evaluation word can only be reflected when it is actually used to describe the evaluation object. For example, in the sentences "marketing profits are large" and "sales risks are large," the word "large" is generally a positive word. However, when used to modify "marketing profits," it is positive, and when used to modify "sales risks," it is negative. Therefore, the sentiment of an evaluation word can only be determined when it is actually combined with the evaluation object. This combination of an evaluation word and an evaluation object is called an evaluation unit. Evaluation unit extraction gathers subjective opinions about a particular review object in a review text by extracting, combining, and organizing the text. Currently, research on extracting evaluation units from product reviews is progressing rapidly. However, due to limitations in the domain knowledge of Chinese finance and open datasets, extracting evaluation units from Chinese financial texts still faces significant challenges.

[0005] The grid model is an advanced method for extracting sentiment evaluation units. It can enumerate possible evaluation objects, evaluation words, and the relationships between them in the text, and adopts an end-to-end approach, which makes up for the error propagation caused by the previous deep learning model's separate extraction of the relationship between evaluation objects and evaluation words.

[0006] When extracting evaluation units from Chinese financial texts, the existing grid model has the following shortcomings:

[0007] (1) The relationship between Chinese characters in evaluation elements is not fully utilized. In Chinese financial texts, there are complex dependencies between evaluation elements. For example, multiple evaluation units can appear in the same text, and each evaluation unit includes two evaluation elements: evaluation words and evaluation objects. Evaluation units also share common elements. Making full use of the dependency between evaluation elements is conducive to improving the extraction effect of evaluation units. However, the existing grid model only uses syntactic dependency information in word meaning representation and does not fully characterize the relationship between candidate evaluation elements when extracting evaluation elements.

[0008] (2) Grid element representation does not distinguish between different types of evaluation elements. Existing grid models assume that regardless of whether a character in a text serves as an evaluation object or an evaluation word, its representation in the network is the same. Therefore, they are not distinguished, and the representation of evaluation words and evaluation objects is not differentiated. However, in Chinese financial commentary texts, the semantics of individual Chinese characters are not clear enough. Chinese characters that are both evaluation words and evaluation objects are common, and their semantics should be differentiated. In addition, due to the large difference in length between evaluation objects and evaluation words, the way their local information is portrayed should be different.

[0009] (3) Insufficient use of the relationships between Chinese characters within evaluation elements. In Chinese financial texts, there is a strong correlation between Chinese characters within evaluation elements. The current grid model uses Chinese characters as the input granularity. With the help of pre-trained language models, it can effectively learn the semantic information of Chinese characters. However, it ignores the correlation between Chinese characters within evaluation elements, resulting in the model being unable to fully learn the semantics of evaluation elements. Summary of the Invention

[0010] To this end, an embodiment of the present invention proposes a Chinese financial evaluation unit extraction method based on inter-character relationships to solve the problems of insufficient utilization of Chinese character relationships between evaluation elements in the existing technology, failure to distinguish different evaluation element types when representing grid elements, and insufficient utilization of Chinese character relationships within evaluation elements.

[0011] According to an embodiment of the present invention, a method for extracting Chinese financial evaluation units based on inter-character relationships includes:

[0012] Encode Chinese financial text sentences in the encoding layer to obtain word vectors;

[0013] Based on syntactic dependencies, we construct an inter-character relationship matrix, feed the character vectors and the inter-character relationship matrix into a graph convolutional neural network model, evaluate the character vectors, enhance the relationship between Chinese characters, and encode the dependencies.

[0014] Construct a relationship grid for each word pair in the evaluation text, and differentiate the words in the rows and columns of the grid through two unit convolution kernels, so that the same word has different encodings when it serves as different emotional evaluation factors;

[0015] Convolution operations are performed using image convolution kernels of different sizes in both row and column directions to enhance the relationship between Chinese characters within the emotional evaluation elements.

[0016] For the elements in the grid, the corresponding word encodings in the rows and columns and the dependency encodings between the two words are integrated, and the labels of the elements are determined by decoding to complete the extraction of evaluation units.

[0017] The Chinese financial evaluation unit extraction method based on inter-character relationships according to an embodiment of the present invention has the following beneficial effects:

[0018] (1) To address the problem of insufficient learning of the Chinese character relationships between evaluation factors, the present invention uses a graph convolutional neural network to learn syntactic dependencies, obtain the dependencies between the components of Chinese financial texts, and embed the trained syntactic adjacency matrix into the grid model, so that the grid model can explicitly utilize the syntactic structure and enhance the learning of the Chinese character relationships between various evaluation factors;

[0019] (2) To address the problem that existing grid models do not distinguish between different types of evaluation elements, the present invention uses two different unit convolution kernels to transform the word vectors on the rows and columns before splicing them into the embedding vectors of the grid elements, so that they respectively indicate that the word comes from the evaluation object or the evaluation word, thereby differentiating the evaluation word and the evaluation object;

[0020] (3) To address the problem of insufficient learning of Chinese character relationships within evaluation elements, the present invention adopts a flat image convolution structure, which can learn the local relationships between adjacent characters and reduce the influence of the text domain on the pre-trained language model. In addition, image convolution kernels of different lengths are used in the row and column directions of the grid model to model the Chinese character relationships within the evaluation words and evaluation objects respectively, which can further differentiate the evaluation words and evaluation objects.

[0021] In addition, the Chinese financial evaluation unit extraction method based on inter-character relationship according to the above embodiment of the present invention may also have the following additional technical features:

[0022] Furthermore, based on the syntactic dependency relationship, an inter-character relationship matrix is ​​constructed. The character vectors and the inter-character relationship matrix are fed into the graph convolutional neural network model. The character vectors are evaluated to enhance the relationship between Chinese characters and the dependency relationship is encoded. The LAGCN model is used to learn and evaluate the relationship between Chinese characters, and the LAGCN model satisfies the following conditional formula:

[0023]

[0024]

[0025]

[0026] Where l is the number of LAGCN layers, represents the word vector of the i-th word in the l-th layer of LAGCN, represents the word vector of the jth word in the (l-1)th layer of LAGCN, σ represents the activation function, N represents the total number of Chinese characters, j represents the jth Chinese character, represents the adjacency strength between the i-th word and the j-th word in the l-th layer of LAGCN, represents the weight matrix, b D represents the bias matrix, r i，j Indicates the dependency relationship encoding between the i-th word and the j-th word, p j Indicates the part-of-speech code of the word where the i-th word is located, d i，j ∈{0, 1}, d i，j = 0, it means that there is no syntactic association between the i-th word and the j-th word, d i，j =1, it means that there is a syntactic association between the i-th word and the j-th word. exp is an exponential function with the natural constant e as the base. sum means adding each dimension of the vector. represents the element corresponding to the i-th row and j-th column in the adjacency tensor, k represents the k-th Chinese character, d i,k ∈{0, 1}, d i,k = 0, it means that there is no syntactic association between the i-th word and the k-th word, d i,k =1, it means that there is a syntactic association between the i-th word and the k-th word. Represents the element corresponding to the i-th row and k-th column in the adjacency tensor, represents the weight matrix, and || represents the vector connection operation.

[0027] Furthermore, in the step of constructing the relationship grid of each word pair in the evaluation text and differentiating the words in the rows and columns of the grid through two unit convolution kernels, the element g in the grid G i,j The following conditions are met:

[0028]

[0029] h′=conv′(h D )

[0030] h″=conv″(h D )

[0031] Among them, W G represents the weight matrix, is the last layer in LAGCN conv′ and conv″ represent convolution functions, h′ represents the word vector after the convolution operation of the convolution function conv′, and h″ represents the word vector after the convolution operation of the convolution function conv″. D Based on Combine to get a vector set.

[0032] Furthermore, the steps of performing convolution operations on the image using different-sized convolution kernels in the row and column directions to enhance the relationship between Chinese characters in the sentiment evaluation factors specifically include:

[0033] The image convolution kernel is used to apply multiple convolution kernels of different sizes to the grid G ​​in both row and column directions, and convolution operations are performed on adjacent words to obtain the network G for decoding. R , G R The calculation formula is as follows:

[0034] G row,2 =conv row,2 (G)

[0035] G col,2 =conv col，2 (G row,2 )

[0036] G row,3 =conv row,3 (G col,2 )

[0037] G R =conv Col,3 (G row,3 )

[0038] Among them, conv row,2 and conv row，3 Represents the convolution function using 1×2 and 1×3 convolution kernels, conv col,2 and conv Col，3 Represents the convolution function using 2×1 and 3×1 convolution kernels, G row,2 , G col,2 , G row,3 An intermediate amount.

[0039] Furthermore, for the elements in the grid, the corresponding word encodings in the rows and columns and the dependency encodings between the two words are integrated, and the label prediction probability distribution of the elements is inferred by decoding to complete the step of extracting the evaluation unit. The label prediction probability distribution of the elements is inferred using the following formula:

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] Among them, W R 、 represents the weight matrix, represents the bias matrix, Indicates G R The element in the i-th row and j-th column, softmax represents the normalized exponential function, maxpooling is the maximum pooling function, The feature vector representing the word pair relationship between the i-th word and the j-th word in the t-1th round, for The initial value of represents the set of predicted probability distributions of all word pairs for the i-th word in the t-1th round, represents the set of predicted probability distributions of all word pairs for the jth word in the t-1th round, Indicates the predicted probability value of the word pair relationship between the i-th word and the j-th word in the t-1th round, for The initial value of represents the middle value, The feature vector representing the relationship between the i-th word and the j-th word in the t-th round, Represents the predicted probability value of the word pair relationship between the i-th word and the j-th word in the t-th round. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and / or additional aspects and advantages of the embodiments of the present invention will become apparent and readily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0047] Figure 1 1 is a flow chart of a method for extracting Chinese financial evaluation units based on inter-character relationships according to an embodiment of the present invention;

[0048] Figure 2 This is an exemplary diagram of constructing an adjacency matrix. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0050] See also Figure 1 The embodiment of the present invention proposes a method for extracting Chinese financial evaluation units based on inter-character relationships, comprising the following steps:

[0051] Step 1: Encode the Chinese financial text sentences in the encoding layer to obtain word vectors.

[0052] The encoding layer uses ERNIE for encoding on a per-character basis. Assume there are N Chinese characters in a financial text sentence c, and ERNIE pre-training yields a set of character vectors h.

[0053] Step 2: Construct an inter-character relationship matrix based on syntactic dependencies, feed the character vectors and the inter-character relationship matrix into the graph convolutional neural network model, enhance the Chinese character relationship between evaluation elements of the character vectors, and encode the dependency relationship.

[0054] The present invention constructs an adjacency matrix through dependency relationships. The initial adjacency matrix is ​​expanded from the dependency relationships between words parsed by an external parser. An exemplary adjacency matrix construction diagram is shown in FIG. Figure 2 As shown, the relationship type between two characters is the dependency type between the words they belong to. The relationship type between characters belonging to the same word is "self." When there is no dependency relationship between the words, the relationship type between characters is "O." HANLP 2.0 was used as the dependency analysis tool. To distinguish different relationship types, each relationship type is encoded with the same dimension as the character vector and randomly initialized.

[0055] The word vector h and the word relationship matrix are fed into the graph convolutional neural network (GCN) model to obtain the trained adjacency tensor A and the vector set h of the learned syntax. D .

[0056] In this embodiment, the graph convolutional neural network model adopts the LAGCN model to learn the relationship between Chinese characters between evaluation factors.

[0057] Specifically, in step 2, the LAGCN model is used to learn the relationship between Chinese characters in the evaluation factors, and the LAGCN model satisfies the following conditional formula:

[0058]

[0059]

[0060]

[0061] Where l is the number of LAGCN layers, represents the word vector of the i-th word in the l-th layer of LAGCN, represents the word vector of the jth word in the (l-1)th layer of LAGCN, σ represents the activation function, N represents the total number of Chinese characters, j represents the jth Chinese character, represents the adjacency strength between the i-th word and the j-th word in the l-th layer of LAGCN, represents the weight matrix, b D represents the bias matrix, r i，j Indicates the dependency relationship encoding between the i-th word and the j-th word, p j Indicates the part-of-speech code of the word where the i-th word is located, d i，j ∈{0, 1}, d i，j = 0, it means that there is no syntactic association between the i-th word and the j-th word, d i，j =1, it means that there is a syntactic association between the i-th word and the j-th word. exp is an exponential function with the natural constant e as the base. sum means adding each dimension of the vector. represents the element corresponding to the i-th row and j-th column in the adjacency tensor, k represents the k-th Chinese character, d i,k ∈{0, 1}, d i,k = 0, it means that there is no syntactic association between the i-th word and the k-th word, d i,k =1, it means that there is a syntactic association between the i-th word and the k-th word. Represents the element corresponding to the i-th row and k-th column in the adjacency tensor, represents the weight matrix, and || represents the vector connection operation.

[0062] Step 3: Construct a relationship grid for each word pair in the evaluation text, and differentiate the words in the rows and columns of the grid through two unit convolution kernels, so that the same word has different encodings when it serves as different emotional evaluation elements.

[0063] In step 3, the element g in the grid G ​​is i，j The following conditions are met:

[0064]

[0065] h′=conv′(h D )

[0066] h″=conv″(h D )

[0067] Among them, W G represents the weight matrix, is the last layer in LAGCN Specifically, it represents the element corresponding to the i-th row and j-th column in the adjacency tensor, conv′ and conv″ represent convolution functions, the sliding step size is 1, h′ represents the word vector after the convolution operation of the convolution function conv′, and h″ represents the word vector after the convolution operation of the convolution function conv″. D Based on Combine to get a vector set.

[0068] Step 4: Use image convolution kernels of different sizes to perform convolution operations in the row and column directions to enhance the relationship between Chinese characters in the emotional evaluation factors.

[0069] Among them, step 4 specifically includes:

[0070] The image convolution kernel is used to apply multiple convolution kernels of different sizes to the grid G ​​in both row and column directions, and convolution operations are performed on adjacent words to obtain the network G for decoding. R , G R The calculation formula is as follows:

[0071] G row,2 =conv row，2 (G)

[0072] G col，2 =conv col，2 (G row，2 )

[0073] G row,3 =conv row,3 (G col,2 )

[0074] G R =conv col,3 (G row,3 )

[0075] Among them, conv row，2 and conv row，3 Represents the convolution function using 1×2 and 1×3 convolution kernels, conv col，2 and conv col，3 Represents the convolution function using 2×1 and 3×1 convolution kernels, G row，2 , G col，2 , G row，3 An intermediate amount.

[0076] Step 5: For the elements in the grid, the corresponding word encodings in the rows and columns and the dependency encodings between the two words are integrated, and the labels of the elements are determined by decoding to complete the extraction of the evaluation unit.

[0077] Among them, in English data, in order to deal with the problem of non-login words, words are usually divided into multiple tokens, and the encoding of the initial token is used as the final encoding of the word. Similarly, after fully considering the relationship between words, the present invention regards the encoding of the first letter of the word as the final encoding of the word and participates in the label prediction work, which speeds up the prediction speed and reduces the inconsistency problem caused by too many labels. The present invention retains the Inference module of GTS, taking into account certain constraints between labels, and through multiple rounds of inference, allows the distribution of word-pair labels in the previous round of global scope to affect the inference of word-pair labels in the next round.

[0078] Specifically, in step 5, the following formula is used to infer the element's label prediction probability distribution:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] Among them, W R 、 represents the weight matrix, represents the bias matrix, Indicates G R The element in the i-th row and j-th column, softmax represents the normalized exponential function, maxpooling is the maximum pooling function, The feature vector representing the word pair relationship between the i-th word and the j-th word in the t-1th round, for The initial value of represents the set of predicted probability distributions of all word pairs for the i-th word in the t-1th round, represents the set of predicted probability distributions of all word pairs for the jth word in the t-1th round, Indicates the predicted probability value of the word pair relationship between the i-th word and the j-th word in the t-1th round, for The initial value of represents the middle value, The feature vector representing the relationship between the i-th word and the j-th word in the t-th round, Represents the predicted probability value of the word pair relationship between the i-th word and the j-th word in the t-th round.

[0085] In summary, the Chinese financial evaluation unit extraction method based on inter-character relationships provided by the present invention has the following beneficial effects:

[0086] (1) To address the problem of insufficient learning of the Chinese character relationships between evaluation factors, the present invention uses a graph convolutional neural network to learn syntactic dependencies, obtain the dependencies between the components of Chinese financial texts, and embed the trained syntactic adjacency matrix into the grid model, so that the grid model can explicitly utilize the syntactic structure and enhance the learning of the Chinese character relationships between various evaluation factors;

[0087] (2) To address the problem that existing grid models do not distinguish between different types of evaluation elements, the present invention uses two different unit convolution kernels to transform the word vectors on the rows and columns before splicing them into the embedding vectors of the grid elements, so that they respectively indicate that the word comes from the evaluation object or the evaluation word, thereby differentiating the evaluation word and the evaluation object;

[0088] (3) To address the problem of insufficient learning of Chinese character relationships within evaluation elements, the present invention adopts a flat image convolution structure, which can learn the local relationships between adjacent characters and reduce the influence of the text domain on the pre-trained language model. In addition, image convolution kernels of different lengths are used in the row and column directions of the grid model to model the Chinese character relationships within the evaluation words and evaluation objects respectively, which can further differentiate the evaluation words and evaluation objects.

[0089] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0090] Although embodiments of the present invention have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A Chinese financial evaluation unit extraction method based on inter-character relationship, characterized in that: The following steps are involved: Encode Chinese financial text sentences in the encoding layer to obtain word vectors; Based on syntactic dependencies, a character relationship matrix is ​​constructed, and the character vectors and the character relationship matrix are fed into a graph convolutional neural network model to enhance the character relationships between evaluation elements and encode the dependencies. The graph convolutional neural network model uses a LAGCN model to learn and evaluate the relationships between Chinese characters. Construct a grid of word pairs in the evaluation text, and differentiate the words in the rows and columns of the grid through two unit convolution kernels, so that the same word has different encodings when it serves as different emotional evaluation factors; among them, the element g in the grid G i，j The following conditions are met: h′=conv′(h D ) h″=conv″(h D ) Among them, W G represents the weight matrix, is the last layer in the LAGCN model Represents the element corresponding to the i-th row and j-th column in the adjacent tensor, conv′ and conv″ represent convolution functions, h′ represents the word vector after the convolution operation of the convolution function conv′, and h″ represents the word vector after the convolution operation of the convolution function conv″. D Based on The vector set is obtained by combining, where l is the number of LAGCN layers, Represents the word vector of the i-th word in the l-th layer of the LAGCN model; Convolution operations are performed using image convolution kernels of different sizes in both row and column directions to enhance the relationship between Chinese characters within the emotional evaluation elements. For the elements in the grid, the corresponding word encodings in the rows and columns and the dependency encodings between the two words are integrated, and the label prediction probability distribution of the elements is inferred by decoding to complete the evaluation unit extraction. Among them, the label prediction probability distribution of the elements is inferred using the following formula: Among them, W R 、 represents the weight matrix, represents the bias matrix, Indicates G R The element in the i-th row and j-th column, softmax represents the normalized exponential function, maxpooling is the maximum pooling function, The feature vector representing the word pair relationship between the i-th word and the j-th word in the t-1th round, for The initial value of represents the set of predicted probability distributions of all word pairs for the i-th word in the t-1th round, represents the set of predicted probability distributions of all word pairs for the jth word in the t-1th round, Indicates the predicted probability value of the word pair relationship between the i-th word and the j-th word in the t-1th round, for The initial value of represents the middle value, The feature vector representing the word pair relationship between the i-th word and the j-th word in the t-th round, Represents the predicted probability value of the word pair relationship between the i-th word and the j-th word in the t-th round.

2. The Chinese financial evaluation unit extraction method based on inter-character relationship according to claim 1 is characterized in that: The LAGCN model satisfies the following conditional formula: in, represents the word vector of the jth word in the (l-1)th layer of LAGCN, σ represents the activation function, N represents the total number of Chinese characters, j represents the jth Chinese character, represents the adjacency strength between the i-th word and the j-th word in the l-th layer of the LAGCN model, represents the weight matrix, b D represents the bias matrix, r i，j Indicates the dependency relationship encoding between the i-th word and the j-th word, p j Indicates the part-of-speech code of the word where the i-th word is located, d i，j ∈{0, 1}, d i，j = 0, it means that there is no syntactic association between the i-th word and the j-th word, d i，j =1, it means that there is a syntactic association between the i-th word and the j-th word, exp is an exponential function with the natural constant e as the base, sum means adding each dimension of the vector, k means the k-th Chinese character, d i,k ∈{0, 1}, d i,k = 0, it means that there is no syntactic association between the i-th word and the k-th word, d i,k =1, it means that there is a syntactic association between the i-th word and the k-th word. Represents the element corresponding to the i-th row and k-th column in the adjacency tensor, represents the weight matrix, and || represents the vector connection operation.

3. The Chinese financial evaluation unit extraction method based on inter-character relationship according to claim 2 is characterized in that: The steps of performing convolution operations on the image using different-sized convolution kernels in the row and column directions to enhance the relationship between Chinese characters in the sentiment evaluation factors include: The image convolution kernel is used to apply multiple convolution kernels of different sizes to the grid G ​​in both row and column directions, and convolution operations are performed on adjacent words to obtain the network G for decoding. R , G R The calculation formula is as follows: G row,2 =conv row,2 (G) G col,2 =conv col,2 (G row,2 ) G row，3 =conv row,3 (G col,2 ) G R =conv col，3 (G row，3 ) Among them, conv row，2 and conv row，3 Represents the convolution function using 1×2 and 1×3 convolution kernels, conv col，2 and conv col，3 Represents the convolution function using 2×1 and 3×1 convolution kernels, G row，2 , G col，2 , G row，3 An intermediate amount.

Images