A text processing method, device, and storage medium

Abstract

The present invention discloses a text processing method, apparatus and storage medium. Among them, the method includes: obtaining a word-level representation and a semantic-level representation of a target text; respectively performing argument component relationship confirmation and argument component classification on the word-level representation and the semantic-level representation, and respectively obtaining an argument component relationship result and an argument component classification result corresponding to the word-level representation and the semantic-level representation, and further respectively obtaining a first argument relationship classification result and a second argument relationship classification result, and obtaining an argument structure of the target text according to the corresponding argument component relationship result, argument component classification result, first argument relationship classification result and second argument relationship classification result. Through the above solution, based on the word-level representation and the semantic-level representation of the text, corresponding argument component classification, argument component relationship confirmation and argument relationship classification can be respectively performed, and then multi-task argument mining can be carried out from different dimensions, which can effectively improve the accuracy of multi-task argument mining.

Description

Technical Field

[0001] This application relates to the technical field of natural language processing, and particularly to a text processing method, apparatus, and storage medium. Background Art

[0002] With the explosive growth of data volume, there is an increasing demand for people to quickly obtain the key content of texts, especially for argument mining, which has received more and more attention.

[0003] During the research and practice of related technologies, the inventors of this application found that in the case of multi-task argument mining, it is not possible to effectively perform argument mining on multi-task data types, resulting in an unsatisfactory effect of argument mining and thus a low accuracy of argument mining. Summary of the Invention

[0004] The main technical problem to be solved by this application is to provide a text processing method, system, and storage medium that can respectively obtain corresponding argument component classifications, argument component relationships, and argument relationship classifications through word-level representation and semantic-level representation, implement multi-task processing of argument mining, obtain corresponding argument structures, and thus improve the accuracy of argument mining.

[0005] To solve the above technical problem, one technical solution adopted by this application is: to provide a text processing method, the method includes: obtaining word-level representation and semantic-level representation of a target text; performing argument component relationship confirmation and argument component classification on the word-level representation to obtain a first argument component relationship result and a first argument component classification result; obtaining a first argument relationship classification result according to the first argument component relationship result and the first argument component classification result; performing argument component relationship confirmation and argument component classification on the semantic-level representation to obtain a second argument component relationship result and a second argument component classification result; obtaining a second argument relationship classification result according to the second argument component relationship result and the second argument component classification result; obtaining the argument structure of the target text according to the first argument component relationship result and the second argument component relationship result, the first argument component classification result and the second argument component classification result, and the first argument relationship classification result and the second argument relationship classification result.

[0006] In one embodiment of the present application, the confirmation of the argument component relationship and the classification of the argument components for the word-level representation are performed to obtain a first argument component relationship result and a first argument component classification result, including: constructing a word-level argument component graph for each of the argument components according to a plurality of argument components corresponding to the preset word-level representation; updating each of the word-level argument component graphs by using a first graph attention network; determining the argument component relationship between the word-level argument component graphs by using a second graph attention network to obtain a first argument component relationship result; and performing argument component classification according to the first argument component relationship result and the updated word-level argument component graphs to obtain the first argument component classification result.

[0007] In one embodiment of the present application, the determining of the argument component relationship between the word-level argument component graphs by using a second graph attention network to obtain a first argument component relationship result includes: determining a first attention coefficient of each word node in the j-th word-level argument component graph for each word node in the i-th word-level argument component graph by using a mutual attention mechanism; and determining a second attention coefficient of each word node in the i-th word-level argument component graph for each word node in the j-th word-level argument component graph by using a mutual attention mechanism; weighting the word nodes in the j-th word-level argument component graph by using the first attention coefficient to obtain a first weighted word node representation; and weighting the word nodes in the i-th word-level argument component graph by using the second attention coefficient to obtain a second weighted word node representation; performing fine-grained alignment on each word node representation in the i-th word-level argument component graph and the first weighted word node representation by using an alignment function to determine a first node corresponding representation corresponding to the i-th word-level argument component graph; performing fine-grained alignment on each word node representation in the j-th word-level argument component graph and the second weighted word node representation by using an alignment function to determine a second node corresponding representation corresponding to the j-th word-level argument component graph; performing a pooling operation on the first node alignment representation and the second node alignment representation respectively to obtain corresponding first relation graph representation and second relation graph representation; predicting the first relation graph representation and the second relation graph representation by using a multi-layer perception mechanism and a bilinear operation to obtain a corresponding argument component relationship prediction probability; and obtaining a first argument component relationship result according to the argument component relationship prediction probability.

[0008] In one embodiment of the present application, the argument component classification is performed according to the first argument component relationship result and the word-level argument component graph before updating to obtain the first argument component classification result, including: determining a first attention coefficient between each word-level argument component graph and the remaining word-level argument component graphs by using a central specific attention mechanism to obtain a neighbor-aware representation corresponding to each word-level argument component graph; and determining a second attention coefficient corresponding to the nodes in each word-level argument component graph by using a soft attention mechanism to obtain a word-level argument component graph representation corresponding to each word-level argument component graph; predicting the word-level argument component graph representation and the neighbor-aware representation by using a multi-layer perception mechanism and a softmax function to obtain a corresponding prediction probability of the argument component type; and obtaining the first argument component classification result according to the prediction probability of the argument component type.

[0009] In one embodiment of the present application, the first argument relationship classification result is obtained according to the first argument component relationship result and the first argument component classification result, including: performing a representation space transformation on the argument components represented at the word level corresponding to the first argument component relationship result and the first argument component classification result by using a multi-layer perception mechanism to obtain a first argument component representation and a second argument component representation; calculating type probabilities for the first argument component representation and the second argument component representation by using a bilinear affine function and a loss function to obtain corresponding probabilities; and obtaining the first argument relationship classification result according to the probabilities.

[0010] In one embodiment of the present application, the confirmation of the argument component relationship and the argument component classification are performed on the semantic-level representation to obtain a second argument component relationship result and a second argument component classification result, including: constructing a semantic-level argument component graph for each argument component according to a plurality of argument components corresponding to the preset semantic-level representation; updating each semantic-level argument component graph by using a first graph attention network; determining the argument component relationship between the semantic-level argument component graphs by using a second graph attention network to obtain the second argument component relationship result; and performing argument component classification according to the second argument component relationship result and the updated semantic-level argument component graph to obtain the second argument component classification result.

[0011] In one embodiment of the present application, the argument component classification is performed according to the second argument component relationship result and the semantic-level argument component diagram before update to obtain the second argument component classification result, including: determining a first attention coefficient between each semantic-level argument component diagram and an adjacent semantic-level argument component diagram by using a central specific attention mechanism to obtain a neighbor perception representation corresponding to each semantic-level argument component diagram; and determining a second attention coefficient corresponding to a node in each semantic-level argument component diagram by using a soft attention mechanism to obtain a semantic-level argument component diagram representation corresponding to each semantic-level argument component diagram; predicting the semantic-level argument component diagram representation and the neighbor perception representation by using a multi-layer perception mechanism and a softmax function to obtain a corresponding prediction probability of the argument component type; and obtaining the second argument component classification result according to the prediction probability of the argument component type.

[0012] In one embodiment of the present application, the second argument relationship classification result is obtained according to the second argument component relationship result and the second argument component classification result, including: performing a representation space transformation on the argument components represented at the semantic level corresponding to the second argument component relationship result and the second argument component classification result by using a multi-layer perception mechanism to obtain a third argument component representation and a fourth argument component representation; calculating type probabilities for the third argument component representation and the fourth argument component representation by using a bilinear affine function and a loss function to obtain corresponding probabilities; and obtaining the second argument relationship classification result according to the probabilities.

[0013] In one embodiment of the present application, the argument structure of the target text is obtained according to the first argument component relationship result, the second argument component relationship result, the first argument component classification result, the second argument component classification result, the first argument relationship classification result, and the second argument relationship classification result, including: calculating an overall prediction probability for the confirmation of the argument component relationship for the first argument component relationship result and the second argument component relationship result by using a mutual learning mechanism and a loss function for the confirmation of the argument component relationship to determine the overall prediction probability for the confirmation of the argument component relationship; calculating an overall prediction probability for the argument component classification for the first argument component classification result and the second argument component classification result by using a mutual learning mechanism and a loss function for the argument component classification to determine the overall prediction probability for the argument component classification; calculating an overall prediction probability for the argument relationship classification for the first argument relationship classification result and the second argument relationship classification result by using a mutual learning mechanism and a loss function for the argument relationship classification to determine the overall prediction probability for the argument relationship classification; and determining the argument structure of the target text according to the overall prediction probability for the confirmation of the argument component relationship, the overall prediction probability for the argument component classification, and the overall prediction probability for the argument relationship classification.

[0014] To solve the above technical problems, another technical solution adopted in this application is: to provide a text processing device, which includes: a memory and a processor coupled to the memory. The memory stores at least one computer program, and when the at least one computer program is loaded and executed by the processor, it is used to implement the above text processing method.

[0015] To solve the above technical problems, yet another technical solution adopted in this application is: to provide a computer-readable storage medium, which stores at least one segment of program, and when the at least one segment of program is loaded and executed by a processor, it is used to implement the above text processing method.

[0016] Different from the prior art, the text processing method provided in this application includes: obtaining a word-level representation and a semantic-level representation of a target text; performing argument component relationship confirmation and argument component classification on the word-level representation to obtain a first argument component relationship result and a first argument component classification result; obtaining a first argument relationship classification result according to the first argument component relationship result and the first argument component classification result; performing argument component relationship confirmation and argument component classification on the semantic-level representation to obtain a second argument component relationship result and a second argument component classification result; obtaining a second argument component relationship classification result according to the second argument component relationship result and the second argument component classification result; obtaining the argument structure of the target text according to the first argument component relationship result and the second argument component relationship result, the first argument component classification result and the second argument component classification result, and the first argument relationship classification result and the second argument relationship classification result; that is, this application respectively obtains corresponding argument component relationships, argument component classifications, and argument relationship classifications through the word-level representation and the semantic-level representation of the target text, and then uses the argument component relationships, argument component classifications, and argument relationship classifications in different dimensions to determine the argument structure corresponding to the target text, realizing multi-task processing of argument mining and improving the accuracy of argument mining. Description of the Drawings

[0017] Figure 1 It is a flowchart of an embodiment of the text processing method of the present invention;

[0018] Figure 2 It is a flowchart of an embodiment of step S2 of the present invention;

[0019] Figure 3 It is a structural diagram of an embodiment of the hierarchical graph neural module in the present invention;

[0020] Figure 4 It is a flowchart of an embodiment of step S23 of the present invention;

[0021] Figure 5It is a flowchart of an embodiment of step S24 of the present invention;

[0022] Figure 6 It is a flowchart of an embodiment of step S3 of the present invention;

[0023] Figure 7 It is a flowchart of an embodiment of step S4 of the present invention;

[0024] Figure 8 It is a flowchart of an embodiment of step S43 of the present invention;

[0025] Figure 9 It is a flowchart of an embodiment of step S44 of the present invention;

[0026] Figure 10 It is a flowchart of an embodiment of step S5 of the present invention;

[0027] Figure 11 It is a flowchart of an embodiment of step S6 of the present invention;

[0028] Figure 12 It is a structural diagram of an embodiment of the text processing device of the present invention;

[0029] Figure 13 It is a structural diagram of an embodiment of the storage medium of the present invention. Detailed Embodiments

[0030] Next, with reference to the accompanying drawings and embodiments, the present invention will be further described in detail. It should be specifically noted that the following embodiments are only used to illustrate the present invention, but do not limit the scope of the present invention. Similarly, the following embodiments are only partial embodiments of the present invention rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0031] Referring to "embodiments" herein means that the specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the present invention. The phrase appears in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0032] Traditional text processing methods, especially for argument mining, include those that classify argument components separately, confirm argument relationships separately, classify argument relationship types separately, and those that only focus on tree structures or only focus on non-tree structures. When dealing with multi-task argument mining, the effect of argument mining is not ideal. Therefore, the argument mining methods in the prior art have poor argument mining effects for multi-tasks, and further result in low accuracy of multi-task argument mining.

[0033] The applicant found in the research that for the situation of multi-task argument mining in the prior art, when conducting argument mining, a unified framework can be established to separately obtain the word-level representation and semantic-level representation of the target text. The corresponding argument component classification, confirmation of argument component relationships, and classification of argument component relationships can be carried out separately in different dimensions, and it is applicable to both tree structures and non-tree structures. Furthermore, multi-dimensional multi-task argument mining can be performed to obtain a more accurate argument structure, effectively improving the accuracy of multi-task argument mining.

[0034] Therefore, a text processing method is proposed. By obtaining the word-level representation and semantic-level representation of the target text; performing confirmation of argument component relationships and argument component classification on the word-level representation to obtain the first argument component relationship result and the first argument component classification result; obtaining the first argument relationship classification result according to the first argument component relationship result and the first argument component classification result; performing confirmation of argument component relationships and argument component classification on the semantic-level representation to obtain the second argument component relationship result and the second argument component classification result; obtaining the second argument component relationship classification result according to the second argument component relationship result and the second argument component classification result; obtaining the argument structure of the target text according to the first argument component relationship result and the second argument component relationship result, the first argument component classification result and the second argument component classification result, and the first argument component relationship classification result and the second argument component relationship classification result.

[0035] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of an embodiment of the text processing method of the present invention; it should be noted that if there are substantially the same results, the method of the present invention is not limited to Figure 1 the process sequence shown, as Figure 1 shown, the method includes the following steps:

[0036] S1. Obtain the word-level representation and semantic-level representation of the target text;

[0037] Among them, the target text is the text that needs to be subjected to argument mining. The word-level representation is the hierarchical division of the words in the text, and the semantic-level representation is the division of the meanings of the sentences in the text.

[0038] Specifically, obtain the target text that needs to be argumentation mined, and perform a hierarchical classification according to the words in the text to represent at the word level, and perform a classification according to the meaning of the sentence to represent at the semantic level.

[0039] In some embodiments, a pre-trained language module can be used to obtain the word-level representation and semantic-level representation of the target text. For example, use the pre-trained language module Bert to obtain the word-level representation H = {h1, …, h n} of the argumentative text, where H ∈ R n×d , and d is the dimension of the output vector of the last layer of the pre-trained language module Bert; and adopt the semantic role labeling technology to learn the semantic role labels of each argumentative component through the pre-trained language module, and determine the semantic-level representation through the semantic role labels. For example: use Bert-SRL to parse each argumentative component and extract a list of semantic role tuples R = {r1, …, r k}, where where and represent the arguments (i.e., the agent and the patient) of the i-th tuple, and VERB i is the predicate (usually a verb).

[0040] S2. Confirm the argumentative component relationships and classify the argumentative components for the word-level representation to obtain the first argumentative component relationship result and the first argumentative component classification result;

[0041] Among them, the confirmation of the argumentative component relationships is to divide the corresponding word-level argumentative component relationships for the word-level representation to obtain the argumentative component relationship result; the classification of the argumentative components is to classify the word-level argumentative components to obtain the argumentative component classification result.

[0042] Specifically, based on the divided word-level representation, confirm its corresponding argumentative component relationships to obtain the corresponding word-level argumentative component relationship result, corresponding to the directed edge prediction task, and perform the classification of the argumentative components to obtain the corresponding word-level argumentative component classification result, corresponding to the node classification task.

[0043] Refer to Figure 2 , Figure 2 which is a schematic flowchart of an embodiment of step S2. Step S2 includes:

[0044] S21. For each argumentative component corresponding to the preset word-level representation, construct a word-level argumentative component graph;

[0045] Among them, each text can be divided into multiple argumentative components. Therefore, the corresponding multiple argumentative components can be determined through the word-level representation of the text division; the word-level argumentative component graph is to learn a fine-grained argumentative component representation from the word level.

[0046] Specifically, the argumentative text is divided into multiple word-level argumentative components through the word-level representation corresponding to the argumentative text. For each word-level argumentative component, the expression of the fine-grained argumentative component is learned from the corresponding word level, and then the corresponding word-level argumentative component graph is obtained.

[0047] In some embodiments, a hierarchical graph neural module can be used to determine multiple argumentative components corresponding to the word-level representation, and construct a word-level argumentative component graph G for each argumentative component. WAC Among them, the co-occurrence relationship describes the relationship of appearing in an argumentative component simultaneously, and this relationship is undirected in the word-level argumentative component graph, and the graph is densely connected. For example: in the word-level argumentative component graph, by taking each word level as a node and representing the co-occurrence relationship between word levels as edges, a word-level argumentative component graph is constructed based on the nodes and edges. Then, the three subtasks of argument mining (i.e., argument component type classification ACTC, argument component relationship confirmation ARI, and argument component relationship type ARTC) are respectively transformed into three tasks of node classification, directed edge prediction, and directed edge type classification in the argument graph.

[0048] See Figure 3 , Figure 3 is the structural block diagram of an embodiment of the hierarchical graph neural module in this application. The hierarchical graph neural module includes: a pre-trained language module Bert, an argumentative component graph attention mechanism, a mutual graph attention mechanism, a multi-layer perception mechanism, a center-specific attention mechanism, a soft attention mechanism, an argumentative component-level attention mechanism, and a bilinear operation; the training process of the hierarchical graph neural model is as follows: the corresponding word-level representation and semantic-level representation of the target text are obtained through the pre-trained language module Bert. According to the word-level representation and semantic-level representation, a word-level argumentative component graph and a semantic-level argumentative component graph are respectively constructed. The argumentative component relationship confirmation of the word-level argumentative component graph and the semantic-level argumentative component graph is respectively performed through the corresponding argumentative component graph attention mechanism and mutual attention mechanism. And the argumentative component classification confirmation of the word-level argumentative component graph and the semantic-level argumentative component graph is respectively performed through the multi-layer perceptron formed by the center-specific attention mechanism, the soft attention mechanism, and the argumentative component-level attention mechanism. Then, the argumentative relationship classification is respectively performed on the argumentative component relationship confirmation results and argumentative component classification confirmation results corresponding to the word-level argumentative component graph and the semantic-level argumentative component graph through the bilinear operation, and the corresponding argumentative relationship classification results are obtained. Thus, the mutual learning mechanism is used to perform mutual learning on the argumentative component relationship confirmation results, argumentative component classification confirmation results, and argumentative relationship classification results corresponding to the word-level argumentative component graph and the semantic-level argumentative component graph respectively, and finally the argumentative structure of the target text is determined.

[0049] In some embodiments, the graph embedding representation can be learned from the word view through a hierarchical graph neural module, and the word view corresponds to the word-level argument component graph. The word view is set as a two-layer graph structure. The first layer is the argument graph, where each argument component is a graph node, and the argument components correspond to the argumentation components. The purpose of this argument graph is to learn the knowledge among the argumentation components from the input text. The second layer is the word-level or semantic-level argument component graph, where each word or semantic role is a graph node, and the fine-grained knowledge within each argument component is learned from the word-level or semantic-level. Specifically, the word-level argument component graph learns the fine-grained argument component representation from the word-level, while the semantic-level argument component graph learns the fine-grained argument component representation from the semantic-level.

[0050] In some embodiments, argumentative texts are collected as training corpora, and the argumentative structures are manually labeled. An argumentative passage is represented as a text sequence P = {w1, …, w n} consisting of n words, and this passage is divided into m argument components X = {x1, …, x m}, where x1 represents the span of the i-th argument component including the starting word index and the ending word index Formally, for each argument component i, its type label (i.e., claim and premise) is predicted in the argument component type classification task; the directed edge label (i.e., relevant or irrelevant) from argument component i to argument component j is predicted in the argument relation confirmation task; the type label (i.e., support or oppose) of the directed edge from argument component i to argument component j is predicted in the argument relation type classification task, and the label values are all from a finite set.

[0051] In some embodiments, the node embedding representation in the word-level argument component graph is initialized with the context word representation H learned from BERT and can be expressed as where, d is the size of the embedding dimension, and represent the context representation and the number of nodes of the p-th node in the word-level argument component graph i, respectively.

[0052] S22. Update each word-level argument component graph using the first graph attention network;

[0053] Among them, the first graph attention network is used to update the representation of word nodes, so the word-level argument component graph can be updated.

[0054] Specifically, the representation of each word-level point in the word-level argument component graph is updated through the first graph attention network. Since the target text has multiple argument components and corresponds to multiple word-level argument component graphs, each word-level argument component graph is updated.

[0055] In some embodiments, the first graph attention network can be a word-level argument component graph attention mechanism. After updating the representation of each word node through this word-level argument component graph attention mechanism, the internal structure of the word-level argument component graph can be constructed by aggregating the neighbor information on the word-level argument component graph. For example, a pair of word nodes (p, q) connected by an undirected edge is set, and the corresponding attention score β pq is calculated. The attention score β pq represents the importance score of the word node q for the representation of the learning node p:

[0056]

[0057] where W WAC ∈ R d×d represents the shared linear transformation in the word-level argument component graph i; represents the representation of the p-th word node; att WAC (·) represents the attention function for the word node.

[0058] To compare the coefficients of different nodes, the softmax function is used to normalize all the attention scores β of the selectable word nodes q pq as follows:

[0059]

[0060] where σ(·) represents the LeakyReLU activation function, N p represents the neighbor of the learning node p, and | | represents the concatenation operation; a WAC ∈ R 2d is an attention vector.

[0061] After obtaining the normalized attention scores, the updated node representation of the learning node p in the argument component graph i is calculated through an update calculation method

[0062]

[0063] where W WAC ∈ R d×d represents the shared linear transformation in the word-level argument component graph i; represents the representation of the p-th word node; represents the normalized attention score; σ(·) represents the LeakyReLU activation function.

[0064] S23. Use the second graph attention network to determine the argument component relationship between the word-level argument component graphs, and obtain the first argument component relationship result;

[0065] Among them, the second graph attention network is used for the argumentation relationship recognition task, that is, to determine the relationship between each word-level argumentation component.

[0066] Specifically, based on the word-level argumentation component graph, the relationship between each pair of argumentation components can be determined by the second graph attention network, corresponding to the directed edge prediction problem, that is, whether there is a directed edge between two argumentation components, corresponding to the directed edge prediction task.

[0067] In some embodiments, the second graph attention network can be a mutual graph attention mechanism. The cross-graph propagation based on the mutual graph attention mechanism is used for the argumentation component relationship determination task, that is, the argumentation relationship recognition task. The argumentation relationship recognition task can be transformed into an edge prediction problem in the argumentation component graph. Therefore, it is necessary to obtain the relationship between each pair of argumentation components.

[0068] Refer to Figure 4 , Figure 4 is a schematic flowchart of an embodiment of step S23. Step S23 includes:

[0069] S231. Use the mutual attention mechanism to determine the first attention coefficient of each word node in the j-th word-level argumentation component graph for each word node in the i-th word-level argumentation component graph; and use the mutual attention mechanism to determine the second attention coefficient of each word node in the i-th word-level argumentation component graph for each word node in the j-th word-level argumentation component graph;

[0070] Wherein, each word level in the target text corresponds to a word node, and there is a word node representation for each word node. Therefore, the word-level argumentation component graph contains multiple word node representations.

[0071] Specifically, first determine all the word node representations corresponding to the word-level argumentation component graphs in the argumentation component pair; that is, determine all the corresponding word node representations in the word-level argumentation component graphs corresponding to each argumentation component pair in the target text. For example: the argumentation component pair includes the i-th argumentation component and the j-th argumentation component, corresponding to the i-th word-level argumentation component graph and the j-th word-level argumentation component graph, determine and represent the representations of all word nodes in the i-th and j-th word-level argumentation component graphs, where and represent the number of nodes in the corresponding word-level argumentation component graphs, and then use the mutual graph attention mechanism to learn the importance score, that is, the attention coefficient, of each semantic node in one semantic-level argumentation component graph for each semantic node in the other semantic-level argumentation component graph.

[0072] In some embodiments, the word-level argument component pair's word-level argument component graph can be the i-th word-level argument component graph and the j-th word-level argument component graph, represented by the word node corresponding to the i-th word-level argument component graph as a query, and represented by the word node corresponding to the j-th word-level argument component graph as the key and value, and then through the second graph attention network, calculate the first attention coefficient of the q-th word node corresponding to the j-th word-level argument component graph with respect to the p-th word node corresponding to the i-th word-level argument component graph, that is, use the graph attention function att ARI ∈R d ×R d →R to calculate the first attention coefficient

[0073]

[0074] where, W p ∈R d×d and W q ∈R d×d are parameter matrices.

[0075] and represented by the word node corresponding to the j-th word-level argument component graph as a query, and represented by the word node corresponding to the i-th word-level argument component graph as the key and value, and then through the second graph attention network, calculate the second attention coefficient of the p-th word node corresponding to the i-th word-level argument component graph with respect to the q-th word node corresponding to the j-th word-level argument component graph, that is, use the graph attention function att ARI ∈R d ×R d →R to calculate the second attention coefficient

[0076]

[0077] where, W p ∈R d×d and W q ∈R d×d are parameter matrices.

[0078] S232. Weight the word nodes in the j-th word-level argument component graph using the first attention coefficient to obtain the first weighted word node representation; and weight the word nodes in the i-th word-level argument component graph using the second attention coefficient to obtain the second weighted word node representation;

[0079] where, weighting the word nodes is to obtain a specific node representation of the word-level argument component graph.

[0080] Specifically, use the first attention coefficient Weight the word nodes in the word-level argument component diagram of the j-th to obtain a specific node representation as i.e., corresponding to the first weighted word node representation, and each element in the first weighted word node representation is calculated as follows:

[0081]

[0082] where is each element in the first weighted node representation, is the first attention coefficient, is the representation of each word-level node in the word-level argument component diagram of the j-th.

[0083] And use the second attention coefficient to weight the word nodes in the word-level argument component diagram of the i-th to obtain a specific node representation as i.e., corresponding to the second weighted word node representation; the calculation method of each element in the second weighted word node representation is as follows:

[0084]

[0085] where is each element in the first weighted node representation, is the first attention coefficient, is the representation of each word-level node in the word-level argument component diagram of the i-th.

[0086] In some embodiments, the softmax function can also be used to normalize the first attention coefficient and the second attention coefficient and then perform weighting to obtain the corresponding weighted word node representation.

[0087] S233. Use the alignment function to perform fine-grained alignment on the representation of each word node in the word-level argument component diagram of the i-th and the first weighted word node representation to determine the first node alignment representation corresponding to the word-level argument component diagram of the i-th;

[0088] where the alignment function is to set the alignment method for the text in the specified device environment.

[0089] Specifically, apply the alignment function to perform fine-grained alignment between the word nodes in the word-level argument component diagram of the i-th, and calculate the node fine-grained alignment representation of the word-level argument component diagram of the i-th according to the first weighted word node representation corresponding to the first node alignment representation; where each node alignment representation v i,pThe calculation method is as follows:

[0090]

[0091] Among them, v i,p represents the fine-grained alignment representation formed by the word node p in the i-th word-level argumentation component diagram being affected by the j-th word-level argumentation component diagram. W v ∈R 4d×d is a parameter matrix, and ⊙ represents element-wise multiplication.

[0092] S234. Use the alignment function to perform fine-grained alignment on each word node representation and the second weighted word node representation in the j-th word-level argumentation component diagram to determine the second node alignment representation corresponding to the j-th word-level argumentation component diagram;

[0093] Specifically, apply the alignment function to perform fine-grained alignment between word nodes in the j-th word-level argumentation component diagram, and calculate the node fine-grained alignment representation of the j-th word-level argumentation component diagram corresponding to the second node alignment representation; among them, the calculation method of each node alignment representation v is as follows: j,q The calculation method is as follows:

[0094]

[0095] Among them, v j,q represents the fine-grained alignment representation formed by the word node q in the j-th word-level argumentation component diagram being affected by the i-th word-level argumentation component diagram.

[0096] S235. Perform pooling operations on the first node alignment representation and the second node alignment representation respectively to obtain the corresponding first relation graph representation and second relation graph representation;

[0097] Among them, the first node alignment representation corresponds to the first relation graph representation, and the second node alignment representation corresponds to the second relation graph representation; according to the word node comparison representation of the two word-level argumentation component diagrams, the corresponding directed edge relationship can be obtained, and then the argumentation component relationship corresponding to the word-level argumentation component diagrams can be determined.

[0098] Specifically, by using the average pooling operation on the node alignment vectors corresponding to the node alignment representations of the two word-level component diagrams, the corresponding relation-specific graph representation can be obtained; for example: perform the average pooling operation on the node alignment representation V i of the i-th word-level argumentation component diagram and the node alignment representation V j of the j-th word-level argumentation component diagram respectively to obtain the relation-specific graph representation and corresponding to the first relation graph representation and the second relation graph representation:

[0099]

[0100]

[0101] where, v i,p represents the fine-grained alignment representation formed by the word node p in the i-th word-level argument component diagram being affected by the j-th word-level argument component diagram; v j,q represents the fine-grained alignment representation formed by the word node q in the j-th word-level argument component diagram being affected by the i-th word-level argument component diagram, and represent the number of nodes in the i-th and j-th word-level argument component diagrams respectively.

[0102] S236. Use a multi-layer perception mechanism and bilinear operation to predict the first relational graph representation and the second relational graph representation, and obtain the corresponding argument component relationship prediction probability;

[0103] Among them, the multi-layer perception mechanism can be a multi-layer perceptron (MLP) with a ReLU activation function. Mainly, through the relational graph representation corresponding to the word-level argument component diagram, the directed edge corresponding to the argument component relationship confirmation task can be obtained, and then the prediction probability of the argument component relationship confirmation can be obtained.

[0104] Specifically, in order to capture the bidirectional attribute of each edge in the word-level argument component diagram, use a multi-layer perceptron (MLP) with a ReLU activation function and bilinear operation to and process the graph representation of each argument component pair, and calculate the prediction probability of the argument component relationship confirmation task as:

[0105]

[0106] where, represents the true directed edge of the ARI from the i-th word-level argument component diagram to the j-th word-level argument component diagram. represents the bilinear algorithm, defined as:

[0107]

[0108] where, W g ∈R (d+1)×d is a parameter matrix.

[0109] S237. Obtain the first argument component relationship result according to the argument component relationship prediction probability.

[0110] Among them, determining the argument component relationship prediction probability corresponds to determining the directed edge prediction task, that is, determining the relationship between two argument components.

[0111] Specifically, based on the predicted probability of the argument component relationship, the argument component relationship between word-level argument component graphs is determined to obtain the first argument component relationship result.

[0112] S24. Perform argument component classification based on the first argument component relationship result and the word-level argument component graph before update to obtain the first argument component classification result.

[0113] Among them, obtaining the first argument component relationship result means obtaining the predicted edges and their probabilities in the word-level argument component graph. Combining with the word nodes in the word-level argument component graph, it can be used for the node classification problem, that is, the type classification task of argument components. Therefore, the word-level argument components can be classified according to the obtained predicted edges and their probabilities, and then the first argument component classification result is obtained.

[0114] Refer to Figure 5 , Figure 5 is a schematic flowchart of an embodiment of step S24. Step S24 includes:

[0115] S241. Use the center-specific attention mechanism to determine the first attention coefficient between each word-level argument component graph and the remaining word-level argument component graphs, and obtain the neighbor-aware representation corresponding to each word-level argument component graph.

[0116] Among them, given a central argument component graph, under the guidance of its adjacent argument component graphs, the neighbor-aware representation of the central argument component graph can be learned.

[0117] First, use the fine-grained attention mechanism to learn the graph representation of each argument component graph, and then apply the argument component-level attention mechanism to aggregate information from adjacent argument component graphs along the predicted edges to the central argument component graph. The fine-grained attention mechanism and the argument component-level attention mechanism constitute a hierarchical attention mechanism. The fine-grained attention mechanism includes the center-specific attention mechanism (i.e., the mutual attention mechanism) and the soft attention mechanism.

[0118] Specifically, by using the mutual graph attention mechanism, on using the weighted sum operation, the center-specific node representation of the j-th argument component graph can be obtained where the argument component graph j ∈ N i represents the neighbor argument component graph j of the central argument component graph i. Then, perform average pooling operation on the center-specific node representation U (i,j) to obtain the center-specific graph representation of the central argument component graph i from the neighbor argument component graph j The calculation is as follows:

[0119]

[0120] Among them, j→i represents the information propagation from the i-th argument component diagram to the j-th argument component diagram.

[0121] Learn rich effective information inside the argument components among word nodes through the soft attention mechanism. For example, apply the soft attention mechanism to learn the attention scores of each node in the i-th argument component diagram. Formally, use the nodes of the i-th argument component diagram as the input, and the learning weight η p (1 < p < n i ) is calculated as follows:

[0122]

[0123] where W Ac is the parameter matrix, b is the bias vector; a is an attention vector learned during training.

[0124] After obtaining the importance of each word node, normalize it through the softmax function. Use the learning weight as the coefficient to fuse all node representations in the argument component diagram to obtain the graph representation z i :

[0125]

[0126] where η p is the learning weight.

[0127] Given a central argument component diagram, learn rich effective information inside the argument components among word nodes through the soft attention mechanism, and then combine the importance information inside the argument components and the interaction information between argument component diagrams through the argument component-level attention mechanism. Therefore, the summary information of the argument component diagrams adjacent to each word-level argument component diagram can be aggregated into the central argument component diagram, and then the neighbor-aware representation of the word-level argument component diagram can be determined.

[0128] For example: Given the graph representation z i of the word-level argument component diagram i and the central-specific graph representation Use the edge probability predicted by the argument relationship prediction task to calculate the attention weight γ j→i :

[0129]

[0130] where W ag ∈R d×d represents the shared linear transformation of all nodes. att ag represents the attention mechanism att ag ∈R d ×R d →R.

[0131] Normalize using the softmax function among all neighbor nodes, and the calculation is as follows:

[0132]

[0133] where N i represents the neighborhood of the argument component graph i; σ represents the activation function of LeakyReLU; a ag ∈R 2d is the attention vector learned during training.

[0134] After obtaining the normalized attention coefficients, obtain the neighbor-aware representation c i of the argument component graph by aggregating the information of its neighbors, and the calculation is as follows:

[0135]

[0136] where is the normalized attention coefficient, corresponding to the first attention coefficient, W ag ∈R d×d represents the shared linear transformation of all nodes, is the center-specific graph representation of the central argument component graph i.

[0137] S242. Use the soft attention mechanism to determine the second attention coefficients corresponding to the nodes in each word-level argument component graph, and obtain the word-level argument component graph representation corresponding to each word-level argument component graph;

[0138] Among them, the soft attention mechanism can learn rich effective information inside the argument components among nodes.

[0139] Specifically, by using the mutual graph attention mechanism, the center-specific node representation of the j-th argument component graph can be obtained by using a weighted sum operation on where the argument component graph j ∈ N i represents the neighbor argument component graph j of the central argument component graph i. Then perform an average pooling operation on the center-specific node representation U (i,j) to obtain the center-specific graph representation of the central argument component graph i from the neighbor argument component graph j, and the calculation is as follows:

[0140]

[0141] where j → i represents the information propagation from the i-th argument component graph to the j-th argument component graph.

[0142] Learn the effective information inside the rich argumentation components between word nodes through the soft attention mechanism. For example, apply the soft attention mechanism to learn the attention scores of each node in the i-th argumentation component graph, corresponding to the second attention coefficient. Formally, use the node representation of the i-th argumentation component graph as the input, and the learning weight η of each fine-grained representation is calculated as follows: p (1 < p < n i )

[0143]

[0144] where W AC is the parameter matrix, b is the bias vector; a is an attention vector learned during training.

[0145] After obtaining the importance of each word node, normalize it through the softmax function. Use the learning weight as the coefficient to fuse all node representations in the argumentation component graph to obtain the graph representation z i , corresponding to the word-level argumentation component graph representation, which is calculated as follows:

[0146]

[0147] where η p is the learning weight.

[0148] S243. Use the multi-layer perception mechanism and the softmax function to predict the word-level argumentation component graph representation and the neighbor-aware representation to obtain the corresponding argumentation component type prediction probability;

[0149] Among them, the combined neighbor-aware representation and the graph representation can obtain the comprehensive representation of the word-level argumentation component graph. Through this comprehensive representation of the word-level argumentation component graph, the classification probability of the argumentation component type can be obtained, that is, the node classification task is performed, that is, the argumentation component type classification task is performed.

[0150] Specifically, by combining the neighbor-aware representation c i of the word-level argumentation component graph i and the graph representation z i of all node representations of the word-level argumentation component graph i, obtain the comprehensive representation of the word-level argumentation component graph i. Apply a multi-layer perceptron and the softmax function to the comprehensive representation of the word-level argumentation component graph i for node classification, that is, perform argumentation component type classification, and then obtain the first argumentation component classification result. The argumentation component type classification is calculated as follows:

[0151]

[0152] where, Indicates the predicted type corresponding to the classification of the argument component type of the argument component.

[0153] S244. Obtain the first argument component classification result according to the predicted probability of the argument component type.

[0154] Among them, the predicted probability of the argument component type is the classification probability of the argument component type.

[0155] Specifically, by obtaining the classification probability of the argument component type, the first argument component classification result can be determined, that is, the classification result of the word-level argument component type corresponding to the word-level argument component diagram is obtained.

[0156] S3. Obtain the first argument relationship classification result according to the first argument component relationship result and the first argument component classification result;

[0157] Among them, the target text contains multiple word-level argument components. The classification of the argument component type of the word-level argument component is used to obtain the first argument component classification result, and the relationship of the word-level argument component is determined to obtain the first argument component relationship result. By determining through the classification of the argument component type and the relationship of the argument component, the argument relationship classification between the argument components can be obtained. Therefore, the first argument relationship classification result can be obtained according to the first argument component relationship result and the first argument component classification result.

[0158] See Figure 6 , Figure 6 is a schematic flowchart of an embodiment of step S3. Step S3 includes:

[0159] S31. Use a multi-layer perception mechanism to perform a representation space conversion on the argument components represented at the word level corresponding to the first argument component relationship result and the first argument component classification result, to obtain a first argument component representation and a second argument component representation;

[0160] Among them, the purpose of performing the representation space conversion is to obtain a related specific representation of the two argument component diagrams in the argument component pair.

[0161] Specifically, based on the relationship specific graph representation corresponding to the first argument component relationship result and the comprehensive representation corresponding to the first argument component classification result, determine the overall graph representation corresponding to the two argument component diagrams, corresponding to the first argument component representation and the second argument representation; for example: given a pair of argument component diagrams (i, j), combine the two features of the argument component diagrams to obtain the overall graph representation of the argument component diagrams i and j and

[0162]

[0163] Apply a shared multi-layer perceptron (MLP) layer to transform the overall graph representation into the same representation space, that is and That is, for each argument component in each pair of word-level argument components (i, j), use the multi-layer perceptron mechanism to perform a representation space transformation on the concatenation result of the first argument component relationship result (the pooling result of the node alignment representation: the relationship graph representation) and the first argument component classification result (the neighbor-aware representation), and then obtain the relationship type representation of the i-th argument component and the relationship type representation of the j-th argument component for each pair of word-level argument components (i, j).

[0164] S32. Use a bilinear affine function and a loss function to calculate the type probabilities for the first argument component representation and the second argument component representation, and obtain the corresponding probabilities;

[0165] Among them, the bilinear affine function corresponds to a bilinear operation, and the loss function is determined by the training objective of the argument relationship confirmation task.

[0166] Specifically, for example: use a bilinear affine function to aggregate the features of the nodes i and j of the argument component graph, and apply a softmax function to classify the type of the directed edge from i to j, so as to obtain the classification probability of the argument relationship between the word-level argument components:

[0167]

[0168] Among them, is the classification probability of the argument relationship between the word-level argument components.

[0169] S33. Obtain the first argument relationship classification result according to the probability.

[0170] Among them, the probability is the predicted probability of the argument relationship between the word-level argument components.

[0171] Specifically, the classification of the word-level argument component relationship can be obtained through the predicted probability of the argument relationship between the word-level argument components, so the first argument relationship classification result can be obtained, and the directed edge type classification task is completed.

[0172] S4. Perform argument component relationship confirmation and argument component classification on the semantic-level representation to obtain the second argument component relationship result and the second argument component classification result;

[0173] Among them, the argument component relationship confirmation is to divide the corresponding semantic-level argument component relationship for the semantic-level representation to obtain the argument component relationship result; the argument component classification is to classify the semantic-level argument components to obtain the argument component classification result.

[0174] Specifically, based on the divided semantic level representation, confirm the corresponding argument component relationships, obtain the argument component relationship results at the corresponding semantic level, and classify the argument components to obtain the corresponding semantic level argument component classification results.

[0175] See Figure 7 , Figure 7 which is a schematic flowchart of an embodiment of step S4. Step S4 includes:

[0176] S41. According to the multiple argument components corresponding to the preset semantic level representation, construct a semantic level argument component graph for each argument component;

[0177] Among them, each text can be divided into multiple argument components. Therefore, the corresponding multiple argument components can be determined through the semantic level representation of text division; the semantic level argument component graph is a fine-grained representation of argument components learned from the semantic level.

[0178] Specifically, the argument text is divided into multiple semantic level argument components through the semantic level representation corresponding to the argument text. For each semantic level argument component, learn the expression of fine-grained argument components from the corresponding semantic level, and then obtain the corresponding semantic level argument component graph.

[0179] In some embodiments, a graph neural module can be used to determine the multiple argument components corresponding to the semantic level representation and construct a semantic level argument component graph G for each argument component SAC , where the co-occurrence relationship describes the relationship of appearing in an argument component simultaneously, and this relationship is undirected in the semantic level argument component graph, and the graph is densely connected. For example: in the semantic level argument component graph, by taking each semantic level as a node and representing the co-occurrence relationship between semantic levels as an edge, a semantic level argument component graph is constructed based on the nodes and edges. Then, the three subtasks of argument mining (i.e., argument component type classification ACTC, argument component relationship confirmation ARI, and argument component relationship type ARTC) are respectively transformed into three tasks of node classification, directed edge prediction, and directed edge type classification in the argument graph.

[0180] In some embodiments, a hierarchical graph neural module can be used to learn the embedding representation of the graph from the semantic view, and the semantic view corresponds to the semantic level argument component graph. Set the semantic view as a two-layer graph structure. The first layer is the argument graph, and each argument component is a graph node. The argument component corresponds to the argument component. The purpose of this argument graph is to learn the knowledge between argument components from the input text; the second layer is the semantic level argument component graph, and each semantic role is a graph node respectively, learning the fine-grained knowledge within each argument component from the semantic level. Specifically, the semantic level argument component graph learns the fine-grained representation of argument components from the semantic level.

[0181] In some embodiments, the node embedding representation in the semantic-level argument component graph is initialized with the contextual semantic representation H learned from BERT, which can be expressed as where d is the size of the embedding dimension, and represent the contextual representation and the number of nodes of the p-th node of the semantic-level argument component graph i, respectively.

[0182] S42. Update each semantic-level argument component graph using the first graph attention network;

[0183] Among them, the first graph attention network is used to update the representation of semantic nodes, so the semantic-level argument component graph can be updated.

[0184] Specifically, a semantic role labeling technique is adopted to learn the semantic role labels of each argument component by using BERT-SRL; for example: use BERT-SRL to parse each argument component and extract a list of semantic role tuples R = {r1,..., r k} where and represent the arguments (i.e., the agent and the patient) of the i-th tuple, and VERB i is the predicate (usually a verb). To construct a semantic-level argument component graph for a given argument component, all the arguments and predicates extracted from the argument component are regarded as the semantic-level nodes of the graph. The arguments and predicate nodes that co-occur in a semantic role tuple are connected in the graph. In addition, an undirected edge is added between each pair of nodes with overlapping words. The embedding of the semantic-level nodes is initialized as the contextual word representation learned from BERT. In particular, for an argument or predicate containing multiple words, an average pooling operation is used to obtain the initial semantic node representation. Use to represent the initialization representation of the p-th semantic node of the given semantic argument component graph. A graph attention network is used to update the representation of each semantic node and construct its internal structure by aggregating the information of its neighbors in the semantic-level argument component graph. More specifically, similar to the word-level argument component graph, an attention function is adopted to learn the updated semantic node representation where and represent the updated representation and the number of nodes of the p-th node in the semantic-level argument semantic graph i, respectively.

[0185] For example: Given a pair of semantic nodes (p, q) connected by an undirected edge, calculate the corresponding attention score β pq , and the attention score β pq represents the importance score of the semantic node q for learning the representation of node p:

[0186]

[0187] Among them, W SAC ∈R d×d is represented as the shared linear transformation in the semantic-level argumentation component diagram i; represents the representation of the p-th semantic node; att SAC (·) represents the attention function for the semantic node.

[0188] To compare the coefficients of different nodes, the softmax function is used to normalize the attention scores β of all selectable semantic nodes q pq as follows:

[0189]

[0190] Among them, σ(·) represents the LeakyReLU activation function, N p represents the neighbors of the learning node p, and || represents the concatenation operation; a SAC ∈R 2d is an attention vector.

[0191] After obtaining the normalized attention scores, the updated node representation of the learning node p in the argumentation component diagram i is calculated by updating the calculation method

[0192]

[0193] Among them, W SAC ∈R d×d is represented as the shared linear transformation in the word-level argumentation component diagram i; represents the representation of the p-th word node; represents the normalized attention score; σ(·) represents the LeakyReLU activation function.

[0194] S43. Use the second graph attention network to determine the argumentation component relationship between semantic-level argumentation component diagrams, and obtain the second argumentation component relationship result;

[0195] Among them, the second graph attention network is used for the argumentation relationship recognition task, that is, to determine the relationship between each semantic-level argumentation component.

[0196] Specifically, based on the semantic-level argumentation component diagram, the relationship between each pair of argumentation components can be determined through the second graph attention network, corresponding to the directed edge prediction problem, that is, whether there is a directed edge between two argumentation components.

[0197] In some embodiments, the second graph attention network may be a mutual graph attention mechanism. The cross-graph propagation based on the mutual graph attention mechanism is used for the argument component relationship determination task, that is, the argument relationship recognition task. The argument relationship recognition task can be transformed into an edge prediction problem in the argument component graph. Therefore, it is necessary to obtain the relationship between each pair of semantic-level argument components.

[0198] See Figure 8 , Figure 8 is a schematic flowchart of an embodiment of step S43. Step S43 includes:

[0199] S431. Use the mutual attention mechanism to determine the first attention coefficient of each semantic node in the j-th semantic-level argument component graph for each semantic node in the i-th semantic-level argument component graph; and use the mutual attention mechanism to determine the second attention coefficient of each semantic node in the i-th semantic-level argument component graph for each semantic node in the j-th semantic-level argument component graph;

[0200] Wherein, each semantic object in the target text corresponds to a semantic node, and each semantic node represents a semantic node. Therefore, the semantic-level argument component graph contains multiple semantic node representations.

[0201] Specifically, first determine all semantic node representations of the corresponding semantic-level argument component graphs in the argument component pair; that is, determine all corresponding semantic node representations in the corresponding semantic-level argument component graphs for each argument component pair in the target text. For example: the argument component pair includes the i-th argument component and the j-th argument component, corresponding to the i-th semantic-level argument component graph and the j-th semantic-level argument component graph, determine and represent the representations of all semantic nodes in the i-th and j-th semantic-level argument component graphs, where and represent the number of nodes in the corresponding semantic-level argument component graphs, and then use the mutual graph attention mechanism to learn the importance score, that is, the attention coefficient, of each semantic node in one semantic-level argument component graph for each semantic node in the other semantic-level argument component graph.

[0202] In some embodiments, the semantic-level argument component graphs of the argument component pair may be the i-th semantic-level argument component graph and the j-th semantic-level argument component graph, with the semantic node representation corresponding to the i-th semantic-level argument component graph as the query, and with the semantic node representation corresponding to the j-th semantic-level argument component graph as the key and value, and then calculate the first attention coefficient of the p-th semantic node in the j-th semantic-level argument component graph for the q-th semantic node representation in the i-th semantic-level argument component graph through the second graph attention network, that is, use the graph attention function attARI ∈R d ×R d →R to calculate the first attention coefficient

[0203]

[0204] where, W p ∈R d×d and W q ∈R d×d are parameter matrices.

[0205] and represented by the semantic node corresponding to the j-th semantic-level argument component diagram as a query, and represented by the semantic node corresponding to the i-th semantic-level argument component diagram as the key and value, and then the second attention coefficient represented by the p-th semantic node in the j-th semantic-level argument component diagram for the q-th semantic node in the i-th semantic-level argument component diagram is calculated through the second graph attention network, that is, using the graph attention function att ARI ∈R d ×R d →R to calculate the second attention coefficient

[0206]

[0207] where, W p ∈R d×d and W q ∈R d×d are parameter matrices.

[0208] S432. Weight the semantic nodes in the j-th semantic-level argument component diagram using the first attention coefficient to obtain the first weighted semantic node representation; and weight the semantic nodes in the i-th semantic-level argument component diagram using the second attention coefficient to obtain the second weighted semantic node representation;

[0209] where, weighting the semantic nodes is to obtain a specific node representation of the semantic-level argument component diagram.

[0210] Specifically, use the first attention coefficient to weight the semantic nodes in the j-th semantic-level argument component diagram, and the specific node representation is that is, corresponding to the first weighted semantic node representation, and the calculation method of each element in the first weighted semantic node representation is as follows:

[0211]

[0212] where, For each element in the first weighted semantic node representation, is the first attention coefficient, is the representation of each semantic-level node in the j-th semantic-level argument component diagram.

[0213] And using the second attention coefficient to weight the semantic nodes in the i-th semantic-level argument component diagram to obtain the specific node representation as that is, the corresponding second weighted semantic node representation; the calculation method for each element in the second weighted semantic node representation is as follows:

[0214]

[0215] where, is each element in the first weighted semantic node representation, is the second attention coefficient, is the representation of each semantic-level node in the i-th semantic-level argument component diagram.

[0216] In some embodiments, the softmax function can also be used to normalize the first attention coefficient and the second attention coefficient and then perform weighting to obtain the corresponding weighted semantic node representation.

[0217] S433. Use the alignment function to perform fine-grained alignment on each semantic node representation in the i-th semantic-level argument component diagram and the first weighted semantic node representation to determine the first node alignment representation corresponding to the i-th semantic-level argument component diagram;

[0218] where the alignment function is to set the alignment method for the text in the specified device environment.

[0219] Specifically, apply the alignment function to perform fine-grained alignment between the semantic nodes in the i-th semantic-level argument component diagram, and calculate the node fine-grained alignment representation of the i-th semantic-level argument component diagram corresponding to the first node alignment representation; where the calculation method for each node alignment representation v is as follows: i,p The calculation method is as follows:

[0220]

[0221] where v i,p represents the fine-grained alignment representation formed by the semantic node p in the i-th semantic-level argument component diagram under the influence of the j-th semantic-level argument component diagram. W v ∈R 4d×dis a parameter matrix, and ⊙ represents element-wise multiplication.

[0222] S434. Use the alignment function to perform fine-grained alignment on each semantic node representation and the second weighted semantic node representation in the j-th semantic level argumentation component diagram to determine the second node alignment representation corresponding to the j-th semantic level argumentation component diagram;

[0223] Specifically, apply the alignment function to perform fine-grained alignment between semantic nodes in the j-th semantic level argumentation component diagram, and calculate the node fine-grained alignment representation of the j-th semantic level argumentation component diagram according to the second weighted semantic node representation corresponding to the second node alignment representation; where each node alignment representation v j,q is calculated as follows:

[0224]

[0225] where, v j,q represents the fine-grained alignment representation formed by the influence of the semantic node q in the j-th semantic level argumentation component diagram by the i-th semantic level argumentation component diagram.

[0226] S435. Perform pooling operations on the first node alignment representation and the second node alignment representation respectively to obtain the corresponding first relationship graph representation and second relationship graph representation;

[0227] where, the first node alignment representation corresponds to the first relationship graph representation, and the second node alignment representation corresponds to the second relationship graph representation; according to the semantic node comparison representation of the two semantic level argumentation component diagrams, the corresponding directed edge relationship can be obtained, and then the argumentation component relationship corresponding between the semantic level argumentation component diagrams can be determined.

[0228] Specifically, by using the average pooling operation on the node alignment vectors corresponding to the node alignment representations of the two semantic level component diagrams, the corresponding relationship-specific graph representation can be obtained; for example: perform average pooling operations on the first node alignment representation V i of the i-th semantic level argumentation component diagram and the second node alignment representation V j of the j-th semantic level argumentation component diagram respectively to obtain the relationship-specific graph representation and corresponding to the first relationship graph representation and the second relationship graph representation:

[0229]

[0230]

[0231] where, v i,pRepresents the fine-grained alignment representation formed by the influence of the semantic node p in the i-th semantic-level argumentation component diagram on the j-th semantic-level argumentation component diagram; v j,q Represents the fine-grained alignment representation formed by the influence of the semantic node q in the j-th semantic-level argumentation component diagram on the i-th semantic-level argumentation component diagram, and Represent the number of nodes in the i-th and j-th semantic-level argumentation component diagrams respectively.

[0232] S436. Use the multi-layer perception mechanism and bilinear operation to predict the first relational graph representation and the second relational graph representation to obtain the corresponding argumentation component relationship prediction probability;

[0233] Among them, the multi-layer perception mechanism can be a multi-layer perceptron (MLP) with a ReLU activation function. Mainly, the directed edges corresponding to the relational graph representation of the semantic-level argumentation component diagram can be obtained, and then the prediction probability of the argumentation component relationship confirmation can be obtained.

[0234] Specifically, in order to capture the bidirectional attribute of each edge in the semantic-level argumentation component diagram, the multi-layer perceptron (MLP) with a ReLU activation function and bilinear operation are used to process the graph representation of each pair of argumentation components and to calculate the prediction probability of the argumentation component relationship confirmation as:

[0235]

[0236] Among them, represents the true directed edge of the ARI from the i-th semantic-level argumentation component diagram to the j-th semantic-level argumentation component diagram. represents the bilinear algorithm, defined as:

[0237]

[0238] Among them, W g ∈R (d+1)×d is a parameter matrix.

[0239] S437. Obtain the second argumentation component relationship result according to the argumentation component relationship prediction probability.

[0240] Among them, determining the argumentation component relationship prediction probability corresponds to determining the directed edge prediction task, that is, determining the relationship between two argumentation components.

[0241] Specifically, based on the argumentation component relationship prediction probability, determine the argumentation component relationship between the semantic-level argumentation component diagrams to obtain the semantic-level argumentation component relationship result, that is, the second argumentation component relationship result.

[0242] S44. Classify the argument components according to the second argument component relationship result and the updated semantic-level argument component diagram to obtain the second argument component classification result.

[0243] Among them, obtaining the second argument component relationship result, that is, obtaining the predicted edges and their probabilities in the semantic-level argument component diagram, combined with the semantic nodes in the semantic-level argument component diagram, can be used for the node classification problem, that is, the type classification task of argument components; therefore, the semantic-level argument components can be classified according to the obtained predicted edges and their probabilities, and then the second argument component classification result can be obtained.

[0244] Refer to Figure 9 , Figure 9 is a schematic flowchart of an embodiment of step S44. Step S44 includes:

[0245] S441. Use the center-specific attention mechanism to determine the first attention coefficient between each semantic-level argument component diagram and the remaining semantic-level argument component diagrams, and obtain the neighbor-aware representation corresponding to each semantic-level argument component diagram;

[0246] Among them, given a central argument component diagram, under the guidance of its adjacent argument component diagrams, the neighbor-aware representation of the central argument component diagram can be learned.

[0247] First, use the fine-grained attention mechanism to learn the graph representation of each argument component diagram, and then apply the argument component-level attention mechanism to aggregate information from the adjacent argument component diagrams to the central argument component diagram along the predicted edges. The fine-grained attention mechanism and the argument component-level attention mechanism constitute a hierarchical attention mechanism. The fine-grained attention mechanism includes the center-specific attention mechanism (i.e., the mutual attention mechanism) and the soft attention mechanism.

[0248] Specifically, by using the mutual graph attention mechanism, on using the weighted sum operation, the center-specific node representation of the j-th argument component diagram can be obtained that is, the first weighted semantic node representation, where the argument component diagram j ∈ N i represents the neighbor argument component diagram j of the central argument component diagram i. Then, perform an average pooling operation on the center-specific node representation U (i,j) to obtain the center-specific graph representation of the central argument component diagram i from the neighbor argument component diagram j The calculation is as follows:

[0249]

[0250] where j → i represents the information propagation from the i-th argument component diagram to the j-th argument component diagram.

[0251] Learn the effective information inside the rich argument components between semantic nodes through the soft attention mechanism. For example, apply the soft attention mechanism to learn the attention scores of each node in the $i$-th argument component graph. Formally, take the node representation of the $i$-th argument component graph as the input, and the learning weight $\eta$ of each fine-grained representation p (1 < p < n I ) is calculated as follows:

[0252]

[0253] where $W$ AC is the parameter matrix and $b$ is the bias vector; $a$ is an attention vector learned during training.

[0254] After obtaining the importance of each semantic node, normalize it through the softmax function. Take the learning weight as the coefficient, and fuse all the node representations in the argument component graph to obtain the graph representation $z$ i :

[0255]

[0256] where $\eta$ p is the learning weight.

[0257] Given a central argument component graph, learn the effective information inside the rich argument components between semantic nodes through the soft attention mechanism, and then combine the importance information inside the argument component and the interaction information between the argument component graphs through the argument component-level attention mechanism. Therefore, the summary information of the adjacent argument component graphs of each semantic-level argument component graph can be aggregated into the central argument component graph, and then the neighbor-aware representation of the semantic-level argument component graph can be determined.

[0258] For example: Given the graph representation $z$ i of the semantic-level argument component graph $i$ and the central-specific graph representation and the edge probability predicted by the argument relationship prediction task, calculate the attention weight $\gamma$ j→i :

[0259]

[0260] where $W$ ag $\in \mathbb{R}$ d×d represents the shared linear transformation of all nodes. att ag represents the attention mechanism att ag $\in \mathbb{R}$ d $\times \mathbb{R}$ d $\to \mathbb{R}$.

[0261] Normalize it among all neighbor nodes using the softmax function, and the calculation is as follows:

[0262]

[0263] Among them, N i represents the neighborhood of the argument component graph i; σ represents the activation function of LeakyReLU; a ag ∈R 2d is the attention vector learned during training.

[0264] After obtaining the normalized attention coefficients, by aggregating the information of its neighbors, the neighbor-aware representation c of the argument component graph is obtained i , and the calculation is as follows:

[0265]

[0266] Among them, is the normalized attention coefficient, corresponding to the first attention coefficient, W ag ∈R d×d represents the shared linear transformation of all nodes, is the center-specific graph representation of the central argument component graph i.

[0267] S442. Use the soft attention mechanism to determine the second attention coefficients corresponding to the nodes in each semantic-level argument component graph, and obtain the semantic-level argument component graph representation corresponding to each semantic-level argument component graph;

[0268] Among them, the soft attention mechanism can learn rich and effective information inside the argument components between nodes.

[0269] Specifically, by using the mutual graph attention mechanism, on the weighted sum operation can be used to obtain the center-specific node representation of the j-th argument component graph, where the argument component graph j ∈ N i represents the neighbor argument component graph j of the central argument component graph i. Then, for the center-specific node representation U (i,j) the average pooling operation is used to obtain the center-specific graph representation of the central argument component graph i from the neighbor argument component graph j. The calculation is as follows:

[0270]

[0271] Among them, j→i represents the information propagation from the i-th argument component graph to the j-th argument component graph.

[0272] Learn rich effective information inside the argument components between word nodes through the soft attention mechanism. For example, apply the soft attention mechanism to learn the attention scores of each node in the i-th argument component graph, corresponding to the second attention coefficient. Formally, use the node representation of the i-th argument component graph as the input, and the learning weight η p (1 < p < n i ) is calculated as follows:

[0273]

[0274] where W WC is the parameter matrix, b is the bias vector; a is an attention vector learned during training.

[0275] After obtaining the importance of each semantic node, normalize it through the softmax function. Use the learning weight as the coefficient to fuse all node representations in the argument component graph to obtain the graph representation z i , corresponding to the semantic-level argument component graph representation, which is calculated as follows:

[0276]

[0277] where η p is the learning weight.

[0278] S443. Use the multi-layer perception mechanism and the softmax function to predict the semantic-level argument component graph representation and the neighbor-aware representation to obtain the corresponding prediction probability of the argument component type;

[0279] Among them, the comprehensive representation of the semantic-level argument component graph can be obtained by combining the neighbor-aware representation and the graph representation. Through this comprehensive representation of the semantic-level argument component graph, the classification probability of the argument component type can be obtained, that is, the node classification task is performed, that is, the argument component classification task is performed.

[0280] Specifically, by combining the neighbor-aware representation c i of the semantic-level argument component graph i and the graph representation z i of all node representations of the semantic-level argument component graph i, obtain the comprehensive representation of the semantic-level argument component graph i and obtain the comprehensive representation of the semantic-level argument component graph j through the same method Apply a multi-layer perceptron (MLP) and the softmax function to the comprehensive representation of the semantic-level argument component graph i for node classification, that is, perform argument component classification, and then obtain the second argument component classification result. The argument component classification is calculated as follows:

[0281]

[0282] Among them, represents the predicted type of the argument component i corresponding to the argument component classification.

[0283] S444. Obtain the second argument component classification result according to the predicted probability of the argument component type.

[0284] Among them, the predicted probability of the argument component type is the classification probability of the argument component type.

[0285] Specifically, through the obtained classification probability of the argument component type, the second argument component classification result can be determined, that is, the classification result of the semantic-level argument component type corresponding to the semantic-level argument component graph is obtained.

[0286] S5. Obtain the second argument component relationship classification result according to the second argument component relationship result and the second argument component classification result;

[0287] Among them, the target text contains multiple semantic-level argument components. The classification of the argument component types of the semantic-level argument components is used to obtain the second argument component classification result, and the relationship of the semantic-level argument components is determined to obtain the second argument component relationship result. By determining through the classification of the argument component types and the relationship of the argument components, the argument relationship classification between the argument components can be obtained. Therefore, the second argument relationship classification result can be obtained according to the second argument component relationship result and the second argument component classification result.

[0288] Refer to Figure 10 , Figure 10 which is a schematic flowchart of an embodiment of step S5. Step S5 includes:

[0289] S51. Use a multi-layer perception mechanism to perform a representation space transformation on the argument components represented at the semantic level corresponding to the second argument component relationship result and the second argument component classification result, and obtain a third argument component representation and a fourth argument component representation;

[0290] Among them, the purpose of performing the representation space transformation is to obtain a related specific representation of the two argument component graphs in the argument component pair.

[0291] Specifically, based on the relationship-specific graph representation corresponding to the second argument component relationship result and the comprehensive representation corresponding to the second argument component classification result, the overall graph representation corresponding to the two argument component graphs is determined, corresponding to the third argument component representation and the fourth argument representation; for example: given a pair of argument component graphs (i, j), the overall graph representations of the argument component graph i and the argument component graph j are obtained by combining the two features of the argument component graph and

[0292]

[0293] Among them, is the comprehensive representation of the semantic-level argument component diagram i, is the comprehensive representation of the semantic-level argument component diagram j, is the node alignment representation V of the i-th semantic-level argument component diagram i corresponding relationship-specific graph representation, is the node alignment representation V of the j-th semantic-level argument component diagram j corresponding relationship-specific graph representation.

[0294] Apply a shared multi-layer perceptron layer (MLP) to transform the overall graph representation into the same representation space, that is and

[0295] S52. Use a bilinear affine function and a loss function to calculate the type probabilities of the third argument component representation and the fourth argument component representation, and obtain the corresponding probabilities;

[0296] Among them, the bilinear affine function corresponds to a bilinear operation, and the loss function is determined by the training objective of the argument relationship confirmation task.

[0297] Specifically, for example: use a bilinear affine function to aggregate the features of the nodes i and j of the argument component diagram, and apply a softmax function to classify the directed edge type from i to j, so as to obtain the classification probability of the argument relationship between the semantic-level argument components:

[0298]

[0299] Among them, is the classification probability of the argument relationship between the semantic-level argument components.

[0300] S53. Obtain the second argument relationship classification result according to the probability.

[0301] Among them, the probability is the predicted probability of the argument relationship between the semantic-level argument components.

[0302] Specifically, through the predicted probability of the argument relationship between the semantic-level argument components, the classification of the semantic-level argument component relationship can be obtained, so the second argument relationship classification result can be obtained, and the directed edge type classification task at the semantic level is completed.

[0303] S6. According to the first argument component relationship result and the second argument component relationship result, the first argument component classification result and the second argument component classification result, and the first argument relationship classification result and the second argument relationship classification result, obtain the argument structure of the target text.

[0304] Among them, the first argument component relationship result and the second argument component relationship result can obtain a total argument component relationship result according to the mutual learning mechanism; the first argument component classification result and the second argument component classification result can obtain a total argument component classification result according to the mutual learning mechanism; the first argument relationship classification result and the second argument relationship classification result can obtain a total argument relationship result according to the mutual learning mechanism.

[0305] Refer to Figure 11 , Figure 11 is a schematic flowchart of an embodiment of step S6. Step S6 includes:

[0306] S61. Use the mutual learning mechanism and the loss function for argument component relationship confirmation to calculate the overall prediction probability of the first argument component relationship result and the second argument component relationship result, so as to determine the overall prediction result of argument component relationship confirmation;

[0307] Among them, the mutual learning mechanism can be the Kullback-Leibler (KL) divergence, and the loss function for argument component relationship confirmation is obtained according to the prediction probability corresponding to the first argument component relationship result and the prediction probability corresponding to the second argument component relationship result.

[0308] Specifically, the prediction probabilities corresponding to the first argument component relationship result and the second argument component relationship result are obtained respectively to optimize the training objective, and the prediction probability corresponding to the word-level argument component graph is used to improve the generalization ability of the semantic-level argument component graph for the argument relationship confirmation task. For example: use the Kullback-Leibler (KL) divergence to measure the prediction values of two different-level argument component graphs and matching scores.

[0309] For the semantic-level argument component graph, cross-entropy is applied for edge prediction to optimize the model. Therefore, the training objective for the argument component relationship confirmation task in the semantic-level argument component graph is:

[0310]

[0311] Among them, is the prediction probability of the argument component relationship confirmation task of the semantic-level argument component graph, is the prediction probability of the argument component relationship confirmation task of the word-level argument component graph; is the training objective of the argument component relationship confirmation task in the semantic-level argument component graph.

[0312] Cross-entropy and KL divergence are also applied in the word-level argument component graph to optimize the objective:

[0313]

[0314] Among them, is the training objective of the argument component relationship confirmation task in the semantic-level argument graph.

[0315] Furthermore, the prediction probability of the word-level argument graph is used to improve the generalization ability of the semantic-level argument graph for the argument component relationship confirmation task.

[0316] Combine the training objectives of the word-view argument graph and the semantic-view argument to form the loss function of the final argument component relationship confirmation task:

[0317]

[0318] Among them, L ARI is the loss function of the argument component relationship confirmation task.

[0319] During the inference process, take the average of the prediction probabilities corresponding to the word-view and semantic-view argument graphs as the overall prediction result of the argument component relationship confirmation task:

[0320]

[0321] Among them, is the prediction probability of the argument component relationship confirmation task of the semantic-level argument graph, is the prediction probability of the argument component relationship confirmation task of the word-level argument graph, is the overall prediction result of the argument component relationship confirmation task.

[0322] S62. Use the mutual learning mechanism and the loss function of the argument component classification to calculate the overall prediction probability of the first argument component classification result and the second argument component classification result to determine the overall prediction result of the argument component classification;

[0323] Among them, the mutual learning mechanism can be the Kullback-Leibler (KL) divergence, and the loss function of the argument component classification is obtained according to the prediction probability corresponding to the first argument component classification result and the prediction probability corresponding to the second argument component classification result.

[0324] Specifically, for the semantic-view argument graph, apply cross-entropy to optimize the node classification objective. In addition, use the Kullback-Leibler (KL) divergence and the prediction probability of the word-view argument graph to improve the generalization ability of the semantic-view argument graph for the argument component type classification task.

[0325] The training objective of the argument component type classification task in the semantic argument graph is:

[0326]

[0327] Among them, For the prediction probability of the classification of argument components in the semantic-level argument component graph, For the prediction probability of the classification of argument components in the word-level argument component graph; For the training objective of the classification task of argument components in the semantic-level argument component graph.

[0328] Apply cross-entropy and KL divergence in the word-view argument graph to optimize the objective:

[0329]

[0330] where, For the training objective of the classification task of argument components in the word-level argument component graph.

[0331] Combine the training objectives of the word-view argument graph and the semantic-view argument to form the loss function for the final classification task of argument components:

[0332]

[0333] where L ACTC Is the loss function for the final classification task of argument components.

[0334] During the inference process, take the mean of the prediction probabilities of the word-view and semantic-view argument graphs as the overall prediction result of the argument component classification task:

[0335]

[0336] where, Is the overall prediction probability of the argument component classification, corresponding to the overall prediction result of the argument component classification task.

[0337] S63. Use the mutual learning mechanism and the loss function of the argument relationship classification to calculate the overall prediction probability of the first argument relationship classification result and the second argument relationship classification result, so as to determine the overall prediction result of the argument relationship classification;

[0338] where, the mutual learning mechanism can be the Kullback-Leibler (KL) divergence, and the loss function of the argument relationship classification is obtained according to the prediction probability corresponding to the first argument classification result and the prediction probability corresponding to the second argument classification result.

[0339] Specifically, for the semantic-view argument graph, cross-entropy is used to optimize the classification objective of the directed edge type. In addition, the Kullback-Leibler (KL) divergence and the posterior probability of the word-view argument graph are used to improve the generalization ability of the semantic-view argument graph for the argument relationship classification task.

[0340] The training objective of the semantic-view argument graph in the argument relationship classification task is:

[0341]

[0342] Among them, is the predicted probability of the argumentation relationship classification of the semantic-level argumentation component graph. is the predicted probability of the argumentation relationship classification of the word-level argumentation component graph; is the training objective of the argumentation relationship classification task in the semantic-level argumentation component graph.

[0343] Apply cross-entropy and KL divergence optimization objectives in the word-view argumentation graph:

[0344]

[0345] Among them, is the training objective of the argumentation relationship classification task in the semantic-level argumentation component graph.

[0346] Combine the training objectives of the word-view argumentation graph and the semantic-view argumentation to form the loss function of the final argumentation relationship classification task:

[0347]

[0348] Among them, L ARTC is the loss function of the argumentation relationship classification task.

[0349] During the inference process, take the average of the predicted probabilities of the argumentation relationship classification corresponding to the word-view and semantic-view argumentation graphs as the overall prediction result of the argumentation relationship classification task:

[0350]

[0351] Among them, is the overall prediction result of the argumentation relationship classification task.

[0352] S64. Determine the argumentation structure of the target text according to the overall prediction result confirmed by the argumentation component relationship, the overall prediction result of the argumentation component classification, and the overall prediction result of the argumentation relationship classification.

[0353] Among them, once the argumentation component relationship, the argumentation component classification, and the argumentation relationship classification of the text are determined, the argumentation structure of the text can be determined. Therefore, through the overall prediction results of the argumentation component relationship confirmed by the word-level representation and the semantic-level representation of the target text, the overall prediction result of the argumentation component classification, and the overall prediction result of the argumentation relationship classification, the argumentation structure of the target text can be determined.

[0354] Specifically, in order to better learn each other's complementary information, jointly train these three related subtasks by minimizing the multi-task joint objective function L joint :

[0355] Ljoint = λ1L ARI + λ2L ACTC + λ3L ARTC

[0356] where λ1, λ2, and λ3 are predefined hyperparameters that determine the importance of the three objective functions.

[0357] In some embodiments, based on the gradient descent optimization algorithm or its variants, the classification error is minimized by iteratively optimizing the loss of the model.

[0358] Different from the prior art, in this embodiment, by obtaining the word-level representation and semantic-level representation of the target text; performing argument component relationship confirmation and argument component classification on the word-level representation to obtain the first argument component relationship result and the first argument component classification result; obtaining the first argument relationship classification result according to the first argument component relationship result and the first argument component classification result; performing argument component relationship confirmation and argument component classification on the semantic-level representation to obtain the second argument component relationship result and the second argument component classification result; obtaining the second argument component relationship classification result according to the second argument component relationship result and the second argument component classification result; obtaining the argument structure of the target text according to the first argument component relationship result and the second argument component relationship result, the first argument component classification result and the second argument component classification result, and the first argument relationship classification result and the second argument relationship classification result. That is, based on the word-level representation and semantic-level representation of the target text, the corresponding argument component relationship results, argument component classification results, and argument relationship classification results are obtained respectively, and the argument structure of the target text can be determined from the argument component relationship results, argument component classification results, and argument relationship classification results in different dimensions, improving the accuracy of argument mining.

[0359] Please refer to Figure 12 , Figure 12 which is a schematic structural diagram of an embodiment of the text processing device of the present invention. The system can execute the steps of the above text processing method. For related content, please refer to the detailed description of the above method and will not be elaborated here.

[0360] The text processing system 200 includes: a memory 210 and a processor 220 coupled to the memory. The memory 210 stores at least one computer program, and when at least one computer program is loaded and executed by the processor 220, it is used to implement the above text processing method.

[0361] Please refer to Figure 13 , Figure 12 which is a schematic structural diagram of an embodiment of the computer-readable storage medium of the present invention. The storage medium 300 stores at least one segment of program 310, and when at least one segment of program 310 is loaded and executed by the processor, it is used to implement the above text processing method.

[0362] Through the word-level representation and semantic-level representation of the target text, based on a unified framework, the above method realizes three interacting tasks: argument component classification, argument component relationship confirmation, and argument component relationship classification, improving the accuracy of the argument structure of argumentative texts and the accuracy of argument mining.

[0363] In several embodiments provided by the present invention, 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 merely illustrative. For example, the division of the modules or units is only a logical function division, and there may be other division methods in actual implementation. For example, 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 couplings or direct couplings or communication connections shown or discussed among each other can be indirect couplings or communication connections through some interfaces, devices, or units, and can be in electrical, mechanical, or other forms.

[0364] 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 can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0365] In addition, the functional units in each embodiment of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0366] If the above-mentioned 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 invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs, etc., which can store program codes.

[0367] The above are only the embodiments of the present invention, and thus do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall similarly be included within the patent protection scope of the present invention.

Claims

1. A text processing method, characterized in that, The method includes: Obtaining the word-level representation and semantic-level representation of the target text; Performing argument component relationship confirmation and argument component classification on the word-level representation to obtain a first argument component relationship result and a first argument component classification result; Obtaining a first argument relationship classification result according to the first argument component relationship result and the first argument component classification result; Performing argument component relationship confirmation and argument component classification on the semantic-level representation to obtain a second argument component relationship result and a second argument component classification result; wherein, according to a plurality of argument components corresponding to the preset semantic-level representation, a semantic-level argument component graph is constructed for each argument component; updating each semantic-level argument component graph by using a first graph attention network; determining the argument component relationship between the semantic-level argument component graphs by using a second graph attention network to obtain a second argument component relationship result; determining a first attention coefficient between each semantic-level argument component graph and an adjacent semantic-level argument component graph by using a center-specific attention mechanism to obtain a neighbor-aware representation corresponding to each semantic-level argument component graph; and determining a second attention coefficient corresponding to a node in each semantic-level argument component graph by using a soft attention mechanism to obtain a semantic-level argument component graph representation corresponding to each semantic-level argument component graph; predicting by using a multi-layer perception mechanism and a softmax function the semantic-level argument component graph representation and the neighbor-aware representation to obtain a corresponding argument component type prediction probability; obtaining a second argument component classification result according to the argument component type prediction probability; Obtaining a second argument relationship classification result according to the second argument component relationship result and the second argument component classification result; Obtaining the argument structure of the target text according to the first argument component relationship result and the second argument component relationship result, the first argument component classification result and the second argument component classification result, and the first argument relationship classification result and the second argument relationship classification result.

2. The method according to claim 1, wherein The performing argument component relationship confirmation and argument component classification on the word-level representation to obtain a first argument component relationship result and a first argument component classification result includes: Constructing a word-level argument component graph for each argument component according to a plurality of argument components corresponding to the preset word-level representation; Updating each word-level argument component graph by using a first graph attention network; Determining the argument component relationship between the word-level argument component graphs by using a second graph attention network to obtain a first argument component relationship result; Performing argument component classification according to the first argument component relationship result and the updated word-level argument component graph to obtain the first argument component classification result.

3. The method according to claim 2, wherein The determining the argument component relationship between the word-level argument component graphs by using a second graph attention network to obtain a first argument component relationship result includes: Using the mutual attention mechanism to determine the first attention coefficients of each word node in the j-th word-level argument component graph with respect to each word node in the i-th word-level argument component graph; and using the mutual attention mechanism to determine the second attention coefficients of each word node in the i-th word-level argument component graph with respect to each word node in the j-th word-level argument component graph; Weighting the word nodes in the j-th word-level argument component graph using the first attention coefficients to obtain a first weighted word node representation; and weighting the word nodes in the i-th word-level argument component graph using the second attention coefficients to obtain a second weighted word node representation; Using an alignment function to perform fine-grained alignment on each word node representation in the i-th word-level argument component graph and the first weighted word node representation to determine the first node corresponding representation corresponding to the i-th word-level argument component graph; Using an alignment function to perform fine-grained alignment on each word node representation in the j-th word-level argument component graph and the second weighted word node representation to determine the second node corresponding representation corresponding to the j-th word-level argument component graph; Performing pooling operations on the first node alignment representation and the second node alignment representation respectively to obtain corresponding first relational graph representation and second relational graph representation; Using a multi-layer perception mechanism and a bilinear operation to predict the first relational graph representation and the second relational graph representation to obtain corresponding argument component relationship prediction probabilities; Obtaining a first argument component relationship result according to the argument component relationship prediction probabilities; 4. The method according to claim 2, wherein Performing argument component classification according to the first argument component relationship result and the word-level argument component graph before update to obtain the first argument component classification result, including: Using a center-specific attention mechanism to determine the first attention coefficients between each word-level argument component graph and the remaining word-level argument component graphs to obtain a neighbor-aware representation corresponding to each word-level argument component graph; And using a soft attention mechanism to determine the second attention coefficients corresponding to the nodes in each word-level argument component graph to obtain a word-level argument component graph representation corresponding to each word-level argument component graph; Using a multi-layer perception mechanism and a softmax function to predict the word-level argument component graph representation and the neighbor-aware representation to obtain corresponding argument component type prediction probabilities; Obtaining a first argument component classification result according to the argument component type prediction probabilities; 5. The method according to claim 1, characterized in that, The obtaining of a first argument relationship classification result according to the first argument component relationship result and the first argument component classification result, including: Using a multi-layer perception mechanism to perform a representation space transformation on the argument components represented by the word-level representations corresponding to the first argument component relationship result and the first argument component classification result to obtain a first argument component representation and a second argument component representation; Using a bilinear affine function and a loss function to calculate type probabilities for the first argument component representation and the second argument component representation to obtain corresponding probabilities; Obtaining a first argument relationship classification result according to the probabilities; 6. The method according to claim 1, characterized in that, Obtaining the argumentation structure of the target text according to the first argument component relationship result, the second argument component relationship result, the first argument component classification result, the second argument component classification result, the first argument relationship classification result, and the second argument relationship classification result includes: Calculating the overall prediction probability of the first argument component relationship result and the second argument component relationship result by using a mutual learning mechanism and a loss function for argument component relationship confirmation to determine the overall prediction probability of argument component relationship confirmation; Calculating the overall prediction probability of the first argument component classification result and the second argument component classification result by using a mutual learning mechanism and a loss function for argument component classification to determine the overall prediction probability of argument component classification; Calculating the overall prediction probability of the first argument relationship classification result and the second argument relationship classification result by using a mutual learning mechanism and a loss function for argument component relationship classification to determine the overall prediction probability of argument relationship classification; Determining the argumentation structure of the target text according to the overall prediction probability of argument component relationship confirmation, the overall prediction probability of argument component classification, and the overall prediction probability of argument relationship classification.

7. A text processing device, characterized in that, The text processing device includes: a memory and a processor coupled to the memory. The memory stores at least one computer program, and when the at least one computer program is loaded and executed by the processor, it is used to implement the method according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The storage medium stores at least one segment of program, and when the at least one segment of program is loaded and executed by the processor, it is used to implement the method according to any one of claims 1-6.

Images