Argument mining method and system based on definition, grammar and integrated collaborative attention

Abstract

The present invention proposes a debate mining system based on definition, grammar and integrated collaborative attention, including: an encoder module for learning context information using a bidirectional LSTM; a feature extraction module for extracting features from word vectors using stacked BiLSTMs, a sequential self-attention mechanism and an auxiliary softmax layer as an implicit aid to the vanishing gradient problem; a gradient reversal adjustment module for capturing class-invariant features using a gradient reversal layer; a parallel collaborative attention definition alignment module for discovering different associated features of input text regarding definition criteria of claims and non-claims using a mechanism based on residual fusion coarse-grained parallel collaborative attention; an attention orthogonal projection learning and output module for merging class-invariant features and features of input text regarding definition criteria of claims and non-claims through an attention orthogonal projection layer, and outputting classification results through a softmax activation function; the present invention can efficiently utilize grammar and definitions of claims for claim detection.

Description

Technical Field

[0001] The present invention relates to the field of argument mining data processing technology, and in particular to an argument mining method and system based on definition, grammar and integrated collaborative attention. Background Art

[0002] Argument mining has always been a major challenge in natural language processing. Argument mining primarily involves segmenting argumentative and non-argumentative text, detecting claims, and parsing argument components. Claim detection primarily involves detecting sentences containing claims. Claim detection is fundamental to argument mining and can also be used to detect fake news.

[0003] Claim detection technology has developed rapidly in recent years. The current research on claim detection mainly includes:

[0004] (1) A supervised learning method based on sentiment and word meta-grammar derivation. This method is limited to machine learning methods, but it lays the foundation for future work in this field.

[0005] (2) Unsupervised corpus-wide claim detection method, which derives a claim sentence as a query, then retrieves sentences in the corpus with a greatly increased prior probability of containing the claim related to the topic, and then uses a simple heuristic method to sort the sentences, thus forming an unsupervised corpus-wide claim detection system. However, this system has poor generalization performance.

[0006] (3) Based on the part-of-speech and dependency claim detection method, semantic and latent syntactic features are exploited by combining the language encoder and the context encoder.

[0007] Although existing methods have achieved good results, there is still no method that can efficiently utilize grammar and definitions of claims for claim detection. Summary of the Invention

[0008] The present invention proposes a method and system for argument mining based on definitions, grammar and integrated collaborative attention, which can efficiently use grammar and definitions of claims to perform claim detection.

[0009] The present invention adopts the following technical solutions.

[0010] A definition-based, grammar-integrated, and collaborative attention-based argument mining system for detecting sentences containing assertions, comprising:

[0011] Encoder module: used to learn context information using a bidirectional long short-term memory neural network (LSTM) to extract word vectors of text words with context information;

[0012] Feature extraction module: used to extract features from word embeddings using stacked BiLSTMs, sequential self-attention mechanism, and auxiliary softmax layer: as an implicit aid to the vanishing gradient problem, it improves the accuracy of the network's low-level features;

[0013] Gradient reversal adjustment module: used to capture class-invariant features, i.e., common features shared between classes, using the gradient reversal layer GRL;

[0014] Parallel Co-Attention Definition Alignment Module: This module is used to discover different related features of the input text regarding the definition criteria of claims and non-claims using a coarse-grained parallel co-attention mechanism based on residual fusion;

[0015] Attention orthogonal projection learning and output module: used to merge the class-invariant features and the features of the input text about the claim definition criteria and non-claim definition criteria through the attention orthogonal projection layer, and output the classification results through the softmax activation function.

[0016] A definition-, grammar-, and collaborative-attention-based argument mining method uses a definition-, grammar-, and collaborative-attention-based argument mining system. The method utilizes a combination of feature projection, definition alignment based on collaborative-attention, and gradient reversal to leverage grammatical and semantic information from the context for claim detection.

[0017] In the method, the encoder module constructs a dictionary based on the dataset, using The word embedding method obtains a word embedding matrix. The words in the input text are passed through the word embedding matrix to obtain the word vector of the word. Then, a bidirectional long short-term memory neural network (LSTM) is used to encode the historical context. The word embedding formula is as follows:

[0018]

[0019]

[0020] Where θ is the set of vectors, Δ is the set of all embedding hierarchies and dependencies, and R(x) is the set of random labels unrelated to x. The training loss function L is optimized using Riemannian gradient descent;

[0021] The encoder formula used by the bidirectional LSTM is as follows:

[0022] i t =σ(W (i) x t +I (i) h t-1 +b i ) Formula 3;

[0023] f t =σ(W(f) x t +I (f) h t-1 +b f ) Formula 4;

[0024] o t =σ(W (o) x t +I (o) h 4-1 +b o ) Formula 5;

[0025] u t =tanh(W (n) x t +I (n) h t-1 +b n ) Formula 6;

[0026] c t =i u ⊙u t +f t ⊙c t-1 Formula 7;

[0027] h t =o t ⊙tanh(c t ) Formula 8;

[0028] where x t is the word vector of the tth word in the sentence input at time t, σ represents the sigmoid activation function, W and I represent the weight matrix, and b represents the bias vector. By forgetting and memorizing new information in the cell state of the bidirectional LSTM, useful information for subsequent moment calculations can be transmitted, while useless information is discarded. The hidden layer state h is output at each time step t , where the forget gate f t , Memory Gate i t AND output gate o t The hidden layer state h at the previous moment t-1 and the current input x t Calculate and finally transform the hidden layer units of forward and backward LSTM and Connect as As output, we get two word context information.

[0029] The feature extraction module consists of stacked BiLSTMs, whose hidden units are processed by a sequential self-attention mechanism, and then optimized by an auxiliary softmax layer to implicitly assist in the vanishing gradient problem and improve the accuracy of low-level features of the network.

[0030] The gradient reversal adjustment module receives the vector output by the last BiLSTM layer in the feature extraction module as input and uses the gradient reversal layer GRL to capture the class-invariant features. The forward propagation and backward propagation in the gradient reversal layer are defined by two opposite equations, specifically:

[0031] GRL(x)=x Formula 9;

[0032]

[0033] The parallel collaborative attention definition alignment module discovers different associated features of the input text regarding the definition criteria of claims and non-claims through a mechanism based on residual fusion coarse-grained parallel collaborative attention, including the following steps:

[0034] Step S1: First, two sets of defined criteria are given, C = [C1, C2, ..., C n ] indicates compliance with the criteria defined by the claim, NC = [NC1, NC2, ..., NC m ] indicates compliance with the criteria of the non-assertion definition;

[0035] The encoder module is used to obtain the input text t and the context vector K of the definition criteria set respectively t , K C , K NC The formula for calculating the residual fusion coarse-grained parallel collaborative attention of the input text with respect to the set of claim definition criteria is as follows:

[0036] g t =masked_avg(K t ) Formula 11;

[0037]

[0038] a C =softmax(e c ) Formula XIII;

[0039]

[0040] r C =s C +K t Formula 15;

[0041] Among them, matrices W1, W2 and vectors b, w3 are learnable parameters;

[0042] Calculate K t The mask average of the obtained vector g t Calculate K as the query vector C Additive soft attention on , get vector sC , and the original input text vector K t Add up to get r C ; At the same time, an equivalent symmetric process is used to calculate K C K t Attention, we get vector r t ; vector r C and vector r t After concatenation, the data is input into the residual network to further extract the features of the input text related to the definition criteria of the claim;

[0043] In the coarse-grained parallel collaborative attention mechanism based on residual fusion, the first block of the residual network is a fully connected layer with 20 neurons, and the remaining block consists of a fully connected layer with 5 neurons and a fully connected layer with 20 neurons. The results of the first and last blocks of the residual network are added to obtain the final output vector def claim ;

[0044] Step S2: A symmetric process equivalent to the above is used to calculate the residual fusion coarse-grained parallel collaborative attention of the input text with respect to the non-claim definition criteria set to obtain the output vector def N-claim ;

[0045] vectordef claim and vector def N-claim Appended to the BiLSTM output in the attention orthogonal projection learning and output module, it helps the transient BiLSTM features to have richer grammatical and semantic information, thereby enhancing the classification effect of the attention orthogonal projection learning and output module.

[0046] The attention orthogonal projection learning and output module merges the class-invariant features of the gradient reversal adjustment module and the features of the input text about the claim definition criteria and the non-claim definition criteria obtained by the parallel collaborative attention definition alignment module through the attention orthogonal projection layer, and learns the optimal feature representation; the attention orthogonal projection formula is as follows:

[0047]

[0048] v (s，i)，(r，i) =proj(v (s，i) , v (r，i) ) Formula XVIII;

[0049]

[0050]

[0051]

[0052]

[0053] where v s and v r Represent the feature vectors from the feature extraction module and the dual-path attention feature extraction module respectively; and v r The vertical space eliminates the homogeneity and thus s Projecting onto it will extract the discriminative features from the class-invariant features; v at each time step s Projection to v r In the orthogonal direction, the time distribution layer TDL is used to process the orthogonal results. Then v s Projection to v s，r In the orthogonal direction, the time distribution layer TDL is used to process the orthogonal results. Use the sequence self-attention mechanism to further improve its feature representation; finally use the softmax activation function to normalize the attention vector Output the classification results.

[0054] The attention orthogonal projection learning and output module and gradient reversal adjustment module are trained in parallel with each other, and a sparse classification focus loss function is used. In the classification setting, the label is y and the loss is defined as:

[0055]

[0056] in is a vector representing the approximate probability distribution between the two classes, and γ is the focusing parameter, which essentially reduces the weight of easily classified samples. A higher value means a higher reduction rate of easily classified samples.

[0057] The method also includes a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it utilizes the definition criteria of claims and non-claims to promote the feature construction of the model, utilizes the combination of feature projection, definition alignment based on fused collaborative attention, and gradient reversal, and makes full use of the grammatical and semantic information of the context for claim detection to detect sentences containing claims, or for fake news detection.

[0058] Compared with the existing technology, the present invention and its preferred embodiment can promote the feature construction of the model by utilizing the definition criteria of claims and non-claims, and can fully utilize the grammatical and semantic information of the context for claim detection by combining feature projection, definition alignment based on fused collaborative attention, and gradient reversal. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0060] Attachment Figure 1It is a schematic diagram of the principle of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0062] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0063] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0064] As shown in the figure, a definition-based, grammar-integrated, and collaborative attention-integrated argument mining system can be used to detect sentences containing assertions. The system includes:

[0065] Encoder module: used to learn context information using a bidirectional long short-term memory neural network (LSTM) to extract word vectors of text words with context information;

[0066] Feature extraction module: used to extract features from word embeddings using stacked BiLSTMs, sequential self-attention mechanism, and auxiliary softmax layer: as an implicit aid to the vanishing gradient problem, it improves the accuracy of the network's low-level features;

[0067] Gradient reversal adjustment module: used to capture class-invariant features, i.e., common features shared between classes, using the gradient reversal layer GRL;

[0068] Parallel Co-Attention Definition Alignment Module: This module is used to discover different related features of the input text regarding the definition criteria of claims and non-claims using a coarse-grained parallel co-attention mechanism based on residual fusion;

[0069] Attention orthogonal projection learning and output module: used to merge the class-invariant features and the features of the input text about the claim definition criteria and non-claim definition criteria through the attention orthogonal projection layer, and output the classification results through the softmax activation function.

[0070] A definition-, grammar-, and collaborative-attention-based argument mining method uses a definition-, grammar-, and collaborative-attention-based argument mining system. The method utilizes a combination of feature projection, definition alignment based on collaborative-attention, and gradient reversal to leverage grammatical and semantic information from the context for claim detection.

[0071] In the method, the encoder module constructs a dictionary based on the dataset, using The word embedding method obtains a word embedding matrix. The words in the input text are passed through the word embedding matrix to obtain the word vector of the word. Then, a bidirectional long short-term memory neural network (LSTM) is used to encode the historical context. The word embedding formula is as follows:

[0072]

[0073]

[0074] Where θ is the set of vectors, Δ is the set of all embedding hierarchies and dependencies, and R(x) is the set of random labels unrelated to x. The training loss function L is optimized using Riemannian gradient descent;

[0075] The encoder formula used by the bidirectional LSTM is as follows:

[0076] i t =σ(W (i) x t +I (i) h t-1 +b i ) Formula 3;

[0077] f t =σ(W (f) x t +I (f) h t-1 +b f ) Formula 4;

[0078] o t =σ(W (o) x t +I (o) h 4-1 +b o ) Formula 5;

[0079] u t =tanh(W (n) x t +I (n) h t-1 +b n ) Formula 6;

[0080] c t =i u ⊙ut +f t ⊙c t-1 Formula 7;

[0081] h t =o t ⊙tanh(c t ) Formula 8;

[0082] where x t is the word vector of the tth word in the sentence input at time t, σ represents the sigmoid activation function, W and I represent the weight matrix, and b represents the bias vector. By forgetting and memorizing new information in the cell state of the bidirectional LSTM, useful information for subsequent moment calculations can be transmitted, while useless information is discarded. The hidden layer state h is output at each time step t , where the forget gate f t , Memory Gate i t AND output gate o t The hidden layer state h at the previous moment t-1 and the current input x t Calculate and finally transform the hidden layer units of forward and backward LSTM and Connect as As output, we get two word context information.

[0083] The feature extraction module consists of stacked BiLSTMs, whose hidden units are processed by a sequential self-attention mechanism, and then optimized by an auxiliary softmax layer to implicitly assist in the vanishing gradient problem and improve the accuracy of low-level features of the network.

[0084] The gradient reversal adjustment module receives the vector output by the last BiLSTM layer in the feature extraction module as input and uses the gradient reversal layer GRL to capture the class-invariant features. The forward propagation and backward propagation in the gradient reversal layer are defined by two opposite equations, specifically:

[0085] GRL(x)=x Formula 9;

[0086]

[0087] The parallel collaborative attention definition alignment module discovers different associated features of the input text regarding the definition criteria of claims and non-claims through a mechanism based on residual fusion coarse-grained parallel collaborative attention, including the following steps:

[0088] Step S1: First, two sets of defined criteria are given, C = [C1, C2, ..., C n ] indicates compliance with the criteria defined by the claim, NC = [NC1, NC2, ..., NCm ] indicates compliance with the criteria of the non-assertion definition;

[0089] The encoder module is used to obtain the input text t and the context vector K of the definition criteria set respectively t , K C , K NC The formula for calculating the residual fusion coarse-grained parallel collaborative attention of the input text with respect to the set of claim definition criteria is as follows:

[0090] g t =masked_avg(K t ) Formula 11;

[0091]

[0092] a C =softmax(e C ) Formula XIII;

[0093]

[0094] r C =s C +K t Formula 15;

[0095] Where matrices W1, W2 and vectors b, w3 are learnable parameters;

[0096] Calculate K t The mask average of the vector g t Calculate K as the query vector C Additive soft attention on , get vector s C , and the original input text vector K t Add up to get r C ; At the same time, an equivalent symmetric process is used to calculate K C K t Note that we get vector r t ; vector r C and vector r t After concatenation, the data is input into the residual network to further extract the features of the input text related to the definition criteria of the claim;

[0097] In the coarse-grained parallel collaborative attention mechanism based on residual fusion, the first block of the residual network is a fully connected layer with 20 neurons, and the remaining block consists of a fully connected layer with 5 neurons and a fully connected layer with 20 neurons. The results of the first and last blocks of the residual network are added to obtain the final output vector def claim ;

[0098] Step S2: A symmetric process equivalent to the above is used to calculate the residual fusion coarse-grained parallel collaborative attention of the input text with respect to the non-claim definition criteria set to obtain the output vector def N-claim ;

[0099] vectordef claim and vector def N-claim Appended to the BiLSTM output in the attention orthogonal projection learning and output module, it helps the transient BiLSTM features to have richer grammatical and semantic information, thereby enhancing the classification effect of the attention orthogonal projection learning and output module.

[0100] The attention orthogonal projection learning and output module merges the class-invariant features of the gradient reversal adjustment module and the features of the input text about the claim definition criteria and the non-claim definition criteria obtained by the parallel collaborative attention definition alignment module through the attention orthogonal projection layer, and learns the optimal feature representation; the attention orthogonal projection formula is as follows:

[0101]

[0102] v (s，i)，(r，i) =proj(v (s，i) , v (r，i) ) Formula XVIII;

[0103]

[0104]

[0105]

[0106]

[0107] where v s and v r Represent the feature vectors from the feature extraction module and the dual-path attention feature extraction module respectively; and v r The vertical space eliminates the homogeneity and thus s Projecting onto it will extract the discriminative features from the class-invariant features; v at each time step s Projection to v r In the orthogonal direction, the time distribution layer TDL is used to process the orthogonal results. Then v s Projection to v s，r In the orthogonal direction, the time distribution layer TDL is used to process the orthogonal results. Use the sequence self-attention mechanism to further improve its feature representation; finally use the softmax activation function to normalize the attention vector Output the classification results.

[0108] The attention orthogonal projection learning and output module and gradient reversal adjustment module are trained in parallel with each other, and a sparse classification focus loss function is used. In the classification setting, the label is y and the loss is defined as:

[0109]

[0110] in is a vector representing the approximate probability distribution between the two classes, and γ is the focusing parameter, which essentially reduces the weight of easily classified samples. A higher value means a higher reduction rate of easily classified samples.

[0111] The method also includes a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it utilizes the definition criteria of claims and non-claims to promote the feature construction of the model, utilizes the combination of feature projection, definition alignment based on fused collaborative attention, and gradient reversal, and makes full use of the grammatical and semantic information of the context for claim detection to detect sentences containing claims, or for fake news detection.

[0112] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0113] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0114] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0115] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0116] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A definition-based, grammar-integrated, and collaborative attention-based argument mining system for detecting sentences containing assertions. Its features include: The system comprises: Encoder module: used to learn context information using a bidirectional long short-term memory neural network (LSTM) to extract word vectors of text words with context information; Feature extraction module: used to extract features from word embeddings using stacked BiLSTMs, sequential self-attention mechanism, and auxiliary softmax layer: as an implicit aid to the vanishing gradient problem, it improves the accuracy of the network's low-level features; Gradient reversal adjustment module: used to capture class-invariant features, i.e., common features shared between classes, using the gradient reversal layer GRL; Parallel Co-Attention Definition Alignment Module: This module is used to discover different related features of the input text regarding the definition criteria of claims and non-claims using a coarse-grained parallel co-attention mechanism based on residual fusion; Attention orthogonal projection learning and output module: used to merge the class-invariant features and the features of the input text about the claim definition criteria and non-claim definition criteria through the attention orthogonal projection layer, and output the classification results through the softmax activation function.

2. A method for mining arguments based on definitions, grammar, and the integration of collaborative attention, using the argument mining system based on definitions, grammar, and the integration of collaborative attention as described in claim 1, characterized in that: The proposed method utilizes a combination of feature projection, definition alignment based on fused co-attention, and gradient reversal to exploit the syntactic and semantic information of the context for claim detection; In the method, the encoder module obtains a word embedding matrix based on the dictionary constructed from the dataset; the words of the input text are passed through the word embedding matrix to obtain the word vector of the word, and then a bidirectional long short-term memory neural network (LSTM) is used to encode the historical context; the word embedding formula is as follows: Where θ is the set of vectors, Δ is the set of all embedding hierarchies and dependencies, and R(x) is the set of random labels unrelated to x. The training loss function L is optimized using Riemannian gradient descent. The encoder formula used by the bidirectional LSTM is as follows: i t =σ(W (i) x t +I (i) h t-1 +b i ) Formula 3; f t =σ(W (f) x t +I (f) h t-1 b f ) Formula 4; o t =σ(W (o) x t +I (o) h t-1 +b o ) Formula 5; u t =tanh(W (n) x t +I (n) h t-1 +b n ) Formula 6; c t =i u ⊙u t +f t ⊙c t-1 Formula 7; h t =o t ⊙tanh(c t ) Formula 8; where x t is the word vector of the tth word in the sentence input at time t, σ represents the sigmoid activation function, W and I represent the weight matrix, and b represents the bias vector. By forgetting and memorizing new information in the cell state of the bidirectional LSTM, useful information for subsequent moment calculations can be transmitted, while useless information is discarded. The hidden layer state h is output at each time step t , where the forget gate f t , Memory Gate i t AND output gate o t The hidden layer state h at the previous moment t-1 and the current input x t Calculate and finally transform the hidden layer units of forward and backward LSTM and Connect as As output, we get two word context information.

3. The argument mining method based on definition, grammar and integrated collaborative attention according to claim 2, characterized in that: The feature extraction module consists of stacked BiLSTMs, whose hidden units are processed by a sequential self-attention mechanism, and then optimized by an auxiliary softmax layer to implicitly assist in the vanishing gradient problem and improve the accuracy of low-level features of the network.

4. The argument mining method based on definition, grammar, and collaborative attention according to claim 3, characterized in that: The gradient reversal adjustment module receives the vector output by the last BiLSTM layer in the feature extraction module as input and uses the gradient reversal layer GRL to capture the class-invariant features. The forward propagation and backward propagation in the gradient reversal layer are defined by two opposite equations, specifically: GRL(x)=x Formula 9; 5. The argument mining method based on definition, grammar and integrated collaborative attention according to claim 2, characterized in that: The parallel collaborative attention definition alignment module discovers different associated features of the input text regarding the definition criteria of claims and non-claims through a mechanism based on residual fusion coarse-grained parallel collaborative attention, including the following steps: Step S1: First, two sets of defined criteria are given, C = [C1, C2, ..., C n ] indicates compliance with the criteria defined by the claim, NC = [NC1, NC2, ..., NC m ] indicates compliance with the criteria of the non-assertion definition; The encoder module is used to obtain the input text t and the context vector K of the definition criteria set respectively t , K C , K NC The formula for calculating the residual fusion coarse-grained parallel collaborative attention of the input text with respect to the set of claim definition criteria is as follows: g t =masked_avg(K t ) Formula 11; a C =softmax(e C ) Formula XIII; r C =s C +K t Formula 15; Where matrices W1, W2 and vectors b, w3 are learnable parameters; Calculate K t The mask average of the obtained vector g t Calculate K as the query vector C Additive soft attention on , get vector s C , and the original input text vector K t Add up to get r C ; At the same time, an equivalent symmetric process is used to calculate K C K t Attention, get vector r t ; vector r C and vector r t After concatenation, the data is input into the residual network to further extract the features of the input text related to the definition criteria of the claim; In the mechanism of coarse-grained parallel collaborative attention based on residual fusion, the first block of the residual network is a fully connected layer with 20 neurons, and the remaining block consists of a fully connected layer with 5 neurons and a fully connected layer with 20 neurons; the results of the first and last blocks of the residual network are added to obtain the final output vector def claim ; Step S2: A symmetric process equivalent to the above is used to calculate the residual fusion coarse-grained parallel collaborative attention of the input text with respect to the non-claim definition criteria set to obtain the output vector def N-claim ; vectordef claim and vector def N-claim Appended to the BiLSTM output in the attention orthogonal projection learning and output module, it helps the transient BiLSTM features to have richer grammatical and semantic information, thereby enhancing the classification effect of the attention orthogonal projection learning and output module.

6. The argument mining method based on definition, grammar and integrated collaborative attention according to claim 2, characterized in that: The attention orthogonal projection learning and output module merges the class-invariant features of the gradient reversal adjustment module and the features of the input text about the claim definition criteria and the non-claim definition criteria obtained by the parallel collaborative attention definition alignment module through the attention orthogonal projection layer, and learns the optimal feature representation; the attention orthogonal projection formula is as follows: v (s，i),(r，i) =proj(v (s,i) , v (r，i) ) Formula XVIII; where v s and v r Represent the feature vectors from the feature extraction module and the dual-path attention feature extraction module respectively; and v r The vertical space eliminates the homogeneity and thus s Projecting onto it will extract the discriminative features from the class-invariant features; v at each time step s Projection to v r In the orthogonal direction, and use the time distribution layer TDL to process the orthogonal results; then v s Projection to v s,r In the orthogonal direction, the orthogonal results are processed with the time distribution layer TDL; for each time step Use the sequence self-attention mechanism to further improve its feature representation; finally use the softmax activation function to normalize the attention vector Output the classification results.

7. The argument mining method based on definition, grammar and integrated collaborative attention according to claim 2, characterized in that: The attention orthogonal projection learning is trained in parallel with the output module and the gradient reversal adjustment module, and a sparse classification focus loss function is used. In the classification setting, the label is y and the loss is defined as: in is a vector representing the approximate probability distribution between the two classes, and γ is the focusing parameter, which essentially downweights the easily classified samples; a higher value means a higher downweighting rate of easily classified samples.

8. A computer-readable storage medium for use in the definition-, grammar-, and collaborative attention-based argument mining method of claim 2, characterized in that: The invention comprises a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it utilizes the definition criteria of claims and non-claims to promote the feature construction of the model, utilizes the combination of feature projection, definition alignment based on fusion collaborative attention, and gradient reversal, and makes full use of the grammatical and semantic information of the context for claim detection to detect sentences containing claims, or for false news detection.

Images