CN117454873A - A sarcasm detection method and system based on knowledge-enhanced neural network model - Google Patents

Abstract

The invention discloses a irony detection method and a irony detection system based on a knowledge-enhanced neural network model, comprising the following steps: s1, screening context information highly related to a text to be detected from an external knowledge source and integrating the context information with an original text; s2, word embedding is carried out on the integrated text data by using a pre-training language model RoBERTa, and the weight of the BiLSTM network is memorized in a bidirectional long-short time manner; s3, constructing a coding model consisting of an 8-layer two-way long-short time memory BiLSTM network, embedding a multi-head self-attention mechanism in the two-way long-short time memory BiLSTM network, and capturing long-distance dependency and local semantic features in a text; s4, classifying and training, and improving the model through an optimization algorithm to obtain a final knowledge-reinforced ironic detection model; s5, acquiring a text to be detected, inputting a knowledge-enhanced ironic detection model, and outputting ironic detection results; the invention enhances the ironic understanding of the model, enables the model to effectively capture more complex language modes, and remarkably improves the ironic detection accuracy and robustness.

Description

A sarcasm detection method and system based on knowledge-enhanced neural network model

Technical field

The present invention relates to the technical fields of natural language processing and text mining, and more specifically to a sarcasm detection method and system based on a knowledge-enhanced neural network model.

Background technique

With the popularity of social media and online platforms, the amount of text data (especially user-generated content) has increased exponentially, and these text data often contain sarcasm and innuendo, which is very important for sentiment analysis, public opinion monitoring and natural language understanding. The task poses a challenge, and therefore, sarcasm detection becomes an important research direction in the field of natural language processing.

Early research methods mainly relied on manual extraction of features such as word frequency and emotional vocabulary, as well as traditional machine learning algorithms such as support vector machines (SVM), decision trees and random forests.

With the rapid development of deep learning technology, models based on convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM) have demonstrated superiority in text representation learning, automatic feature extraction, and classification accuracy. performance.

However, most current sarcasm detection models still mainly focus on analyzing the semantic and structural features within the text, while relatively ignoring the possible close relationship between the text and external knowledge sources; to make things more complicated, sarcasm and innuendo are usually It is highly context-dependent and it is often difficult to obtain satisfactory detection performance by relying solely on a single text analysis method.

Therefore, how to effectively integrate external knowledge sources and achieve high accuracy and robustness in sarcasm detection tasks is an urgent problem that needs to be solved by those skilled in the art.

Contents of the invention

In view of this, the present invention provides a sarcasm detection method and system based on a knowledge-enhanced neural network model to solve the problems mentioned in the background art.

In order to achieve the above objects, the present invention adopts the following technical solutions:

A sarcasm detection method based on knowledge-enhanced neural network model, including the following steps:

S1. Filter contextual information that is highly relevant to the text to be detected from external knowledge sources, and integrate the filtered contextual information with the original text;

S2. Use the pre-trained language model RoBERTa to perform word embedding on the integrated text data, and initialize the weights of the bidirectional long short-term memory BiLSTM network;

S3. Construct a coding model consisting of an 8-layer bidirectional long short-term memory BiLSTM network, and embed a multi-head self-attention mechanism in the bidirectional long short-term memory BiLSTM network to capture long-distance dependencies and local semantic features in the text;

S4. Input the word embedding of the pre-trained language model RoBERTa into the encoding model, perform multi-classification training through the fully connected layer and multi-head self-attention mechanism, and the classifier composed of the softmax activation function, and improve the model through the optimization algorithm. Obtain the final knowledge-enhanced sarcasm detection model;

S5. Collect the text to be detected and input the knowledge-enhanced sarcasm detection model to output the sarcasm detection results.

Preferably, the specific content of step S1 is:

S11. Obtain the sentences most relevant to the original text from different external knowledge sources as candidate contexts;

S12. Use the BERTScore text similarity algorithm to sort all candidate contexts, and select the candidate context that best matches the original text as the context of the original text based on the text similarity score;

S13. Connect the extracted context with the original text using the EOS tag indicating the end of the sequence.

Preferably, the BERTScore text similarity algorithm is specifically:

Among them, A and B are the text to be detected and the text in the external knowledge source respectively.

Preferably, the specific steps to obtain the pre-trained language model RoBERTa are:

Collect a large amount of unlabeled text data D = (d ₁ , d ₂ ,...d _n );

Perform word segmentation on each document d _i to obtain word sequence T={t ₁ , t ₂ ,...,t _ni };

Initialize the word embedding matrix E = {e ₁ , e ₂ ,..., e _V }, where e _i is the embedding vector of the i-th word in the vocabulary, and V is the size of the vocabulary;

Based on the Transformer architecture and its self-attention mechanism, the pre-trained language model is trained and the model parameters are optimized using maximum likelihood estimation. After the training is completed, the pre-trained language model RoBERTa is obtained.

Preferably, the specific content of model training in step S4 is:

S41. Divide the integrated text data into a training set and a test set, use the pre-trained language model RoBERTa, and convert the words in the training set into a word embedding matrix X = {x ₁ , x ₂ ,..., x _N } and input it into the coding model middle;

S42. Obtain the hidden state sequence through the 8-layer bidirectional long short-term memory BiLSTM network of the encoding model;

S43. Introduce a multi-head attention mechanism to obtain the attention weight and average weight of each layer;

S44. According to the average weight, input the fully connected layer, and use the activation function softmax to build a classifier to classify the output of the fully connected layer;

S45. Based on the classification results, use the cross-entropy loss function and Adam optimizer to optimize the model;

S46. Update the weight of the model according to the gradient of the loss function to improve the performance of the model.

Preferably, the self-attention mechanism is specifically:

Among them, Q, K, and V are query, key, and value matrices respectively, and d _k is the dimension of the key.

Preferably, step S4 also includes model testing and evaluation: evaluating, analyzing and summarizing the test results of the model through accuracy, recall and F-value, and optimizing and improving model performance.

Preferably, the constructed 8-layer bidirectional long short-term memory BiLSTM network is specifically:

Among them, H _t is the hidden state of the t-th time step, and X _t is the input of the t-th time step;

The attention weight α ^l of each layer is:

Among them, T is the sequence length, is the similarity between the target time step t and the source time step j in the l-th layer BiLSTM;

in, is the hidden state of the target sequence in layer l at time t-1;/> is the hidden state of the source sequence in layer l at time j; a is a learnable function;

average weight for:

The input of the fully connected layer and softmax classifier is expressed as:

The fully connected layer F is:

F(D)＝W _f ·D+b _f

Among them, W _f and b _f are the weights and biases of the fully connected layer respectively;

The classifier C built using the activation function softmax is:

C(F)=soft max(W _c ·F+b _c )

Among them, W _c and b _c are the weight and bias of the classifier respectively;

The cross entropy loss function L is:

Among them, N is the number of labeled samples, y _i is the actual label of the i-th sample, and C(xi ₎ represents the model prediction value of the i-th sample, ranging from [0,1].

Preferably, the rules for optimization using the Adam optimizer are as follows:

m _t =β ₁ m _t-1 +(1-β ₁ )g _t

Among them, m _t and v _t are the estimated values of the first-order moment and the second-order moment respectively, β ₁ and β ₂ are attenuation factors, g _t is the gradient of the loss function L with respect to the model parameter σ, α represents the learning rate, and ε is the prevention factor Divided by a small constant of zero, _σt is the model parameter at time t.

A sarcasm detection system based on a knowledge-enhanced neural network model, based on the sarcasm detection method based on a knowledge-enhanced neural network model, including: a text collection module, a text integration module, a knowledge-enhanced sarcasm detection model and a model construction training module ;

The knowledge-enhanced sarcasm detection model includes the pre-trained language model RoBERTa and the bidirectional long short-term memory BiLSTM model;

Text collection module, used to collect text to be detected and external knowledge sources;

The text integration module is used to filter contextual information that is highly relevant to the text to be detected from external knowledge sources, and integrate the filtered contextual information with the original text;

The pre-trained language model RoBERTa is used to embed words on the integrated text data and initialize the weights of the bidirectional long short-term memory BiLSTM network;

The model construction training module is used to build a coding model composed of an 8-layer bidirectional long short-term memory BiLSTM network. A multi-head self-attention mechanism is embedded in the bidirectional long short-term memory BiLSTM network to capture long-distance dependencies and local semantic features in the text; The word embedding of the pre-trained language model RoBERTa is input into the encoding model, and multi-classification training is performed through the fully connected layer and the multi-head self-attention mechanism, as well as the classifier composed of the softmax activation function, and the model is improved through the optimization algorithm to obtain two-way Long short-term memory BiLSTM model, and the final knowledge-enhanced sarcasm detection model;

The knowledge-enhanced sarcasm detection model is used to input the collected text to be detected and output the sarcasm detection results.

It can be seen from the above technical solutions that compared with the existing technology, the present invention discloses a sarcasm detection method and system based on a knowledge-enhanced neural network model, which combines external knowledge sources to provide contextual information and enhance the model's understanding of sarcasm. making it versatile and more robust than models that do not use external information;

Different from previous work that mainly used simpler neural network architectures, this invention adopts a multi-layer approach, including a pre-trained language model RoBERTa, an 8-layer bidirectional long short-term memory BiLSTM network and a multi-head attention mechanism, so that the model can effectively capture more complex The language model significantly improves the accuracy and robustness of sarcasm detection.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the provided drawings without exerting creative efforts.

Figure 1 is a schematic diagram of the overall framework topology of a sarcasm detection method based on a knowledge-enhanced neural network model provided by the present invention;

Figure 2 is a schematic topology diagram of the pre-trained language model RoBERTa provided by the present invention;

Figure 3 is a schematic topology diagram of the bidirectional long short-term memory BiLSTM network model and the multi-head attention mechanism provided by the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, 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.

The embodiment of the present invention discloses a sarcasm detection method based on a knowledge-enhanced neural network model, which includes the following steps:

S1. Filter contextual information that is highly relevant to the text to be detected from external knowledge sources, and integrate the filtered contextual information with the original text;

S2. Use the pre-trained language model RoBERTa to perform word embedding on the integrated text data, and initialize the weights of the bidirectional long short-term memory BiLSTM network;

S3. Construct a coding model consisting of an 8-layer bidirectional long short-term memory BiLSTM network, and embed a multi-head self-attention mechanism in the bidirectional long short-term memory BiLSTM network to capture long-distance dependencies and local semantic features in the text;

S4. Input the word embedding of the pre-trained language model RoBERTa into the encoding model, perform multi-classification training through the fully connected layer and multi-head self-attention mechanism, and the classifier composed of the softmax activation function, and improve the model through the optimization algorithm. Obtain the final knowledge-enhanced sarcasm detection model;

S5. Collect the text to be detected and input the knowledge-enhanced sarcasm detection model to output the sarcasm detection results.

In order to further implement the above technical solution, the specific content of step S1 is:

S11. Obtain the sentences most relevant to the original text from different external knowledge sources as candidate contexts;

S12. Use the BERTScore text similarity algorithm to sort all candidate contexts, and select the candidate context that best matches the original text as the context of the original text based on the text similarity score;

S13. Connect the extracted context with the original text using the EOS tag indicating the end of the sequence.

In this embodiment, external knowledge sources include Wikipedia, New York Times, and BBC data;

For Wikipedia, find the sentence most relevant to the original text as a candidate context;

For NewYorkTimes, the natural language processing tool spacy is first used to identify named entities in the original text, and the interface NYTAPI is used to retrieve the most relevant sentences as candidate contexts;

For the British Broadcasting Corporation data BBC, the interface GDELTDocAPI is used to search for the top ten headlines most relevant to entities in the original text as candidate contexts.

In order to further implement the above technical solution, the BERTScore text similarity algorithm is specifically:

Among them, A and B are the text to be detected and the text in the external knowledge source respectively.

In order to further implement the above technical solution, the specific steps to obtain the pre-trained language model RoBERTa are:

Collect a large amount of unlabeled text data D = (d ₁ , d ₂ ,...d _n );

Perform word segmentation on each document d _i to obtain word sequence T={t ₁ , t ₂ ,...,t _ni };

Initialize the word embedding matrix E = {e ₁ , e ₂ ,..., e _V }, where e _i is the embedding vector of the i-th word in the vocabulary, and V is the size of the vocabulary;

Based on the Transformer architecture and its self-attention mechanism, the pre-trained language model is trained and the model parameters are optimized using maximum likelihood estimation. After the training is completed, the pre-trained language model RoBERTa is obtained.

In this embodiment, the method of maximum likelihood estimation of MLE optimization model parameters is:

Among them, θ is the model parameter of the pre-trained language model RoBERTa, and N is the size of the training set.

In order to further implement the above technical solution, the specific content of model training in step S4 is:

S41. Divide the integrated text data into a training set and a test set, use the pre-trained language model RoBERTa, and convert the words in the training set into a word embedding matrix X = {x ₁ , x ₂ ,..., x _N } and input it into the coding model middle;

S42. Construct an 8-layer BiLSTM model, which can be expressed as:

Among them, H _t is the hidden state of the t-th time step, and X _t is the input of the t-th time step;

S43. In the training model, a multi-head attention mechanism is introduced to capture key information in the text more effectively:

For each layer l of the 8-layer BiLSTM, there is a hidden state sequence The attention weight α ^l of each layer is:

Among them, T is the sequence length, is the similarity between the target time step t and the source time step j in the l-th layer BiLSTM;

in, is the hidden state of the target sequence in layer l at time t-1;/> is the hidden state of the source sequence in layer l at time j; a is a learnable function;

After obtaining the attention weight α ^l of each layer, the average weight can be calculated

S44. Then add these As the input of the fully connected layer and softmax classifier, it can be expressed as:

In order to further integrate the output of the bidirectional long short-term memory network BiLSTM and prepare the data for use by the classifier, a fully connected layer F is used:

F(D)＝W _f ·D+b _f

Among them, W _f and b _f are the weights and biases of the fully connected layer respectively;

Next, a classifier C built using the activation function softmax is used to classify the output of the fully connected layer:

C(F)=softmax(W _c ·F+b _c )

Among them, W _c and b _c are the weight and bias of the classifier respectively;

S45. In order to quantify the performance of the model and improve the model through optimization algorithms, the cross-entropy loss function L is used:

Among them, N is the number of labeled samples, y _i is the actual label of the i-th sample, and C(xi ₎ represents the model prediction value of the i-th sample, ranging from [0,1];

Optimization uses the Adam optimizer for optimization, and its rules are as follows:

m _t =β ₁ m _t-1 +(1-β ₁ )g _t

Among them, m _t and v _t are the estimated values of the first-order moment and the second-order moment respectively, β ₁ and β ₂ are attenuation factors, usually set to 0.9 and 0.999, g _t is the loss function L about the bidirectional long short-term memory network model parameters The gradient of σ, α represents the learning rate, ε is a small constant to prevent division by zero, usually set to 1×10 ^-8 , and σ _t is the model parameter at time t;

S46. Finally, after each training cycle, update the weight of the model according to the gradient of the loss function to improve the performance of the model.

In order to further implement the above technical solutions, the self-attention mechanism is specifically:

Among them, Q, K, and V are query, key, and value matrices respectively, and d _k is the dimension of the key.

In order to further implement the above technical solution, step S4 also includes model testing and evaluation: the test results of the model are evaluated, analyzed and summarized through accuracy, recall and F-value, and the model performance is optimized and improved.

In this embodiment, the performance index usage accuracy is:

The recall rate is:

The F1 score is:

In this embodiment, contextual information is generated through external knowledge sources (such as Wikipedia, New York Times, and BBC data), and then fused with the original text data; then, the text is converted into numerical values in the word embedding layer The vectors are then processed through the encoding layer and the multi-head attention mechanism layer.

In the encoding layer, a knowledge-enhanced neural network model is used, which can more effectively capture complex patterns in the text; the multi-head attention mechanism layer further extracts key information in the text and uses it in the classification layer, to predict whether a text is sarcastic or not.

In another embodiment, multiple sets of comparative experiments were conducted to verify the effectiveness of the model:

First, on the Semeval-2018Task3 data set, we experimentally explored different network models (including BERT-GRU-Softmax based on the pre-trained language model and the gated recurrent neural network GRU, based on the pre-trained language model and the basic long short-term memory LSTM The performance of BERT-LSTM-Softmax network, three baseline models, and the inventive model) on the sarcasm detection task.

This invention adopts the bidirectional long short-term memory BiLSTM model as shown in Figure 3. The pre-trained RoBERTa neural network model is shown in Figure 2. The Twitter comment text from the Semeval-2018Task3 data set is mainly used to train and evaluate the sarcasm detection model. , the data set was randomly divided into training set, test set and verification set according to the ratio of 8:1:1.

In terms of model construction, the present invention uses a knowledge-enhanced neural network model that obtains contextual information from external knowledge sources (such as Wikipedia, the New York Times, and the British Broadcasting Corporation data BBC). The model mainly consists of an embedding layer and a coding layer. , composed of a multi-head attention mechanism layer and a classification layer. The embedding layer uses the pre-trained RoBERTa model, and the encoding layer uses an 8-layer bidirectional long short-term memory BiLSTM network;

For optimization, use the Adam optimizer and set the weight decay to 1.0e-2; use OneCycleLR as the learning rate scheduler, and set the maximum learning rate to 1.0e-5; set the training cycles to 20, each cycle containing 256 steps; The size of the hidden layer is set to 512, and the dropout rate is set to 0.5; during the training process, the micro-batch size is set to 8, and the maximum training period is set to 20; the storage mode of the word embedding is set to GPU, and 500 is added a preheating step;

Finally, the model uses a softmax layer for classification, outputs the probability that the sample belongs to the two categories of sarcasm or non-sarcasm, and evaluates the performance of the model by comparing it with the true label.

Construct other comparison models and conduct comparative tests. The experimental results on the Semeval2018 data set are as follows:

Table 1 shows:

Different network model methods Marco-F1(%) Pretrained language models and gated recurrent neural networks 79.50 Pre-trained language model and basic long short-term memory network 80.65 Singh et al. (2019) Attention mechanism and emoticon textualization 80.31 Potamias et al. (2020) Converter model and recurrent convolutional network 80.00 Turban et al. (2022) Multi-model ensemble training network 79.80 Invention (Wikipedia) 82.97

Table 1 shows the performance of each model on the Semeval-2018 Task3 data set. From the experimental result data, it can be seen that after introducing Wikipedia context information, the present invention significantly surpassed other baseline models with an F1 score of 82.97%, proving that the model Superiority in sarcasm recognition.

The experimental results using different knowledge sources as context are shown in Table 2:

external knowledge sources Macro-F1(%) No external knowledge is introduced 80.65 New York Times 80.77 BBC data 81.39 Wikipedia 82.97

Table 2 shows the experimental results of different knowledge sources; even if the model proposed by the present invention does not introduce external knowledge, it still reaches an F1 score of 80.65%, which fully proves that by integrating the pre-trained language model RoBERTa, long short-term memory LSTM and attention mechanism , this neural network model has powerful capabilities; when the contextual information of the original text is introduced, all knowledge sources, including Wikipedia, New York Times and BBC data, can improve the performance of the model; specifically, using Wikipedia The model as a knowledge source achieved an F1 score of 82.97%, which is a significant improvement compared to 80.65% without using external knowledge sources; the New York Times and British Broadcasting Corporation data BBC as knowledge sources also respectively made the model's F1 score reach 82.97%. achieved 80.77% and 81.39%; these data clearly show that contextual information from external knowledge sources is very critical in predicting sarcasm.

The experimental results of the model proposed in this invention on different knowledge enhancement methods are shown in Table 3:

Data augmentation strategy Macro-F1(%) back translation 77.58 synonym replacement 80.69 word order replacement 79.21 this invention 82.97

Table 3 shows the experimental results of different data enhancement strategies; looking at the car from it, the three strategies have different effects on the sarcasm detection task. Among them, the synonym replacement strategy has slightly improved, reaching an F1 score of 80.69%, slightly exceeding 80.65% without data enhancement, which shows that synonym replacement, as a data enhancement strategy, can improve the generalization ability of the model to a certain extent; For back translation and word order replacement, the model only achieved F1 scores of 77.58% and 79.21%, which are slightly lower than the model without data augmentation, probably because these two strategies may also introduce data diversity while increasing data diversity. Some noise was added, thus affecting the performance of the model; when the present invention introduced Wikipedia as contextual information into the model, the performance was significantly improved, and the F1 score reached 82.97%, further proving that contextual information in external knowledge sources can provide Richer semantic information helps the model better understand the semantics in the text.

A sarcasm detection system based on a knowledge-enhanced neural network model, based on a sarcasm detection method based on a knowledge-enhanced neural network model, including: a text acquisition module, a text integration module, a knowledge-enhanced sarcasm detection model and a model construction training module;

The knowledge-enhanced sarcasm detection model includes the pre-trained language model RoBERTa and the bidirectional long short-term memory BiLSTM model;

Text collection module, used to collect text to be detected and external knowledge sources;

The text integration module is used to filter contextual information that is highly relevant to the text to be detected from external knowledge sources, and integrate the filtered contextual information with the original text;

The pre-trained language model RoBERTa is used to embed words on the integrated text data and initialize the weights of the bidirectional long short-term memory BiLSTM network;

The model construction training module is used to build a coding model composed of an 8-layer bidirectional long short-term memory BiLSTM network. A multi-head self-attention mechanism is embedded in the bidirectional long short-term memory BiLSTM network to capture long-distance dependencies and local semantic features in the text; The word embedding of the pre-trained language model RoBERTa is input into the encoding model, and multi-classification training is performed through the fully connected layer and the multi-head self-attention mechanism, as well as the classifier composed of the softmax activation function, and the model is improved through the optimization algorithm to obtain two-way Long short-term memory BiLSTM model, and the final knowledge-enhanced sarcasm detection model;

The knowledge-enhanced sarcasm detection model is used to input the collected text to be detected and output the sarcasm detection results.

Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on its differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. As for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant details, please refer to the description in the method section.

The above description of the disclosed embodiments enables those skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be practiced in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. A sarcasm detection method based on a knowledge-enhanced neural network model, which is characterized by including the following steps: S1. Filter contextual information that is highly relevant to the text to be detected from external knowledge sources, and integrate the filtered contextual information with the original text; S2. Use the pre-trained language model RoBERTa to perform word embedding on the integrated text data, and initialize the weights of the bidirectional long short-term memory BiLSTM network; S3. Construct a coding model consisting of an 8-layer bidirectional long short-term memory BiLSTM network, and embed a multi-head self-attention mechanism in the bidirectional long short-term memory BiLSTM network to capture long-distance dependencies and local semantic features in the text; S4. Input the word embedding of the pre-trained language model RoBERTa into the encoding model, perform multi-classification training through the fully connected layer and multi-head self-attention mechanism, and the classifier composed of the softmax activation function, and improve the model through the optimization algorithm. Obtain the final knowledge-enhanced sarcasm detection model; S5. Collect the text to be detected and input the knowledge-enhanced sarcasm detection model to output the sarcasm detection results. 2. A sarcasm detection method based on knowledge-enhanced neural network model according to claim 1, characterized in that the specific content of step S1 is: S11. Obtain the sentences most relevant to the original text from different external knowledge sources as candidate contexts; S12. Use the BERTScore text similarity algorithm to sort all candidate contexts, and select the candidate context that best matches the original text as the context of the original text based on the text similarity score; S13. Connect the extracted context with the original text using the EOS tag indicating the end of the sequence. 3. A sarcasm detection method based on knowledge-enhanced neural network model according to claim 2, characterized in that the BERTScore text similarity algorithm is specifically: Among them, A and B are the text to be detected and the text in the external knowledge source respectively. 4. A sarcasm detection method based on knowledge-enhanced neural network model according to claim 1, characterized in that the specific steps of obtaining the pre-trained language model RoBERTa are: Collect a large amount of unlabeled text data D = (d ₁ , d ₂ ,...d _n ); Perform word segmentation on each document d _i to obtain word sequence T={t ₁ , t ₂ ,...,t _ni }; Initialize the word embedding matrix E = {e ₁ , e ₂ ,..., e _V }, where e _i is the embedding vector of the i-th word in the vocabulary, and V is the size of the vocabulary; Based on the Transformer architecture and its self-attention mechanism, the pre-trained language model is trained and the model parameters are optimized using maximum likelihood estimation. After the training is completed, the pre-trained language model RoBERTa is obtained. 5. A kind of sarcasm detection method based on knowledge-enhanced neural network model according to claim 1, characterized in that the specific content of step S4 model multi-classification training is: S41. Divide the integrated text data into a training set and a test set, use the pre-trained language model RoBERTa, and convert the words in the training set into a word embedding matrix X = {x ₁ , x ₂ ,..., x _N } and input it into the coding model middle; S42. Obtain the hidden state sequence through the 8-layer bidirectional long short-term memory BiLSTM network of the encoding model; S43. Introduce a multi-head attention mechanism to obtain the attention weight and average weight of each layer; S44. According to the average weight, input the fully connected layer, and use the activation function softmax to build a classifier to classify the output of the fully connected layer; S45. Based on the classification results, use the cross-entropy loss function and Adam optimizer to optimize the model; S46. Update the weight of the model according to the gradient of the loss function to improve the performance of the model. 6. A sarcasm detection method based on knowledge-enhanced neural network model according to claim 4 or 5, characterized in that the self-attention mechanism is specifically: Among them, Q, K, and V are query, key, and value matrices respectively, and d _k is the dimension of the key. 7. A sarcasm detection method based on knowledge-enhanced neural network model according to claim 1, characterized in that, after step S4, it also includes model testing and evaluation: the test results of the model pass accuracy, recall and F value. Conduct evaluation, analysis and generalization, and optimize and improve model performance. 8. A sarcasm detection method based on knowledge-enhanced neural network model according to claim 5, characterized in that the constructed 8-layer bidirectional long short-term memory BiLSTM network is specifically: Among them, H _t is the hidden state at the t-th time step, and X _t is the word embedding input at the t-th time step; The attention weight α ^l of each layer is: Among them, T is the length of the word sequence, is the similarity between the target time step t and the source time step j in the l-th layer BiLSTM; in, is the hidden state of the target sequence in layer l at time t-1;/> is the hidden state of the source sequence in layer l at time j; a is a learnable function; average weight for: The input of the fully connected layer and softmax classifier is expressed as: The fully connected layer F is: F(D)＝W _f ·D+b _f Among them, W _f and b _f are the weights and biases of the fully connected layer respectively; The classifier C built using the activation function softmax is: C(F)=softmax(W _c ·F+b _c ) Among them, W _c and b _c are the weight and bias of the classifier respectively; The cross entropy loss function L is: Among them, N is the number of labeled samples, y _i is the actual label of the i-th sample, and C(xi ₎ represents the model prediction value of the i-th sample, ranging from [0,1]. 9. A sarcasm detection method based on knowledge-enhanced neural network model according to claim 8, characterized in that the rules for optimization using the Adam optimizer are specifically: m _t =β ₁ m _t-1 +(1-β ₁ )g _t Among them, m _t and v _t are the estimated values of the first-order moment and the second-order moment respectively, β ₁ and β ₂ are attenuation factors, g _t is the gradient of the loss function L with respect to the model parameter σ, α represents the learning rate, and ε is the prevention factor Divided by a small constant of zero, _σt is the model parameter at time t. 10. A sarcasm detection system based on a knowledge-enhanced neural network model, characterized in that, based on a sarcasm detection method based on a knowledge-enhanced neural network model according to any one of claims 1-9, it includes: a text collection module, Text integration module, knowledge-enhanced sarcasm detection model and model building training module; The knowledge-enhanced sarcasm detection model includes the pre-trained language model RoBERTa and the bidirectional long short-term memory BiLSTM model; Text collection module, used to collect text to be detected and external knowledge sources; The text integration module is used to filter contextual information that is highly relevant to the text to be detected from external knowledge sources, and integrate the filtered contextual information with the original text; The pre-trained language model RoBERTa is used to embed words on the integrated text data and initialize the weights of the bidirectional long short-term memory BiLSTM network; The model construction training module is used to build a coding model composed of an 8-layer bidirectional long short-term memory BiLSTM network. A multi-head self-attention mechanism is embedded in the bidirectional long short-term memory BiLSTM network to capture long-distance dependencies and local semantic features in the text; The word embedding of the pre-trained language model RoBERTa is input into the encoding model, and multi-classification training is performed through the fully connected layer and the multi-head self-attention mechanism, as well as the classifier composed of the softmax activation function, and the model is improved through the optimization algorithm to obtain two-way Long short-term memory BiLSTM model, and the final knowledge-enhanced sarcasm detection model; The knowledge-enhanced sarcasm detection model is used to input the collected text to be detected and output the sarcasm detection results.