Deep learning-based forum bar-sperm speech recognition method and device

Abstract

The invention discloses a forum lever speech recognition method and device based on deep learning. The method comprises the steps of obtaining comment data and performing preprocessing; calculating a text sentiment score based on sentiment words, negative words, adverbs and exclamation marks in the preprocessed comment data; calculating emoticon emotion scores based on emoticons in the comment data; obtaining a final emotion score based on the text emotion score and the expression emotion score; vectorizing and compressing the preprocessed comment data to obtain text sentence vector representation; performing weighted fusion on the text sentence vector representation and the final sentiment score to obtain sentiment value weighted sentence embedding representation; using the sentiment value weighted sentence embedding representation and the corresponding label to train a sentiment classification model; the trained sentiment classification model is used for identifying whether the speech is a bar-sperm speech or not. According to the method, sentiment words, negative words, adverbs, other words, exclamation marks and emoticons are comprehensively considered, and the recognition problem of the lever speech containing hidden contexts is solved.

Description

Technical Field

[0001] The present invention relates to the field of data mining technology, and in particular to a method and device for identifying forum debaters' remarks based on deep learning. Background Art

[0002] Social forums have become an indispensable part of our lives. However, the emergence of controversial speech (i.e., speech that deliberately stirs up controversy) has seriously disrupted the normal social atmosphere and even affected social stability, so it is extremely important to detect such speech. In recent years, controversial speech usually adopts an obscure expression, combining the actual topic context in the comments and using topic keywords to maliciously satirize or ridicule, but traditional detection methods are difficult to effectively identify these implicit emotional expressions. Summary of the invention

[0003] The purpose of the present invention is to address the deficiencies in the prior art and provide a method and device for identifying forum debaters' remarks based on deep learning.

[0004] The objective of the present invention is achieved through the following technical solution: a method for identifying forum trolls' speech based on deep learning, comprising:

[0005] Obtain and preprocess comment data, wherein the comment data includes comments by argumentative people and comments by non-argumentative people;

[0006] Calculate the text sentiment score based on the sentiment words, negative words, adverbs and exclamation marks in the preprocessed comment data;

[0007] Calculate the expression sentiment score based on the emojis in the comment data;

[0008] Based on the text sentiment score and expression sentiment score, the final sentiment score is obtained;

[0009] Vectorize and compress the preprocessed comment data to obtain the text sentence vector representation;

[0010] The text sentence vector representation and the final sentiment score are weighted and fused to obtain the sentiment value-weighted sentence embedding representation;

[0011] The sentiment classification model is trained using the sentence embedding representation weighted by sentiment values ​​and their corresponding labels. The trained sentiment classification model is used to identify whether the speech is from a polemical person.

[0012] Furthermore, the preprocessing includes:

[0013] First, the short texts of the comment data are screened and annotated, then word segmentation is performed, and finally words irrelevant to sentiment analysis are removed.

[0014] Furthermore, the text sentiment score is calculated based on sentiment words, negative words, adverbs and other words and exclamation marks in the pre-processed comment data, including:

[0015] P represents the sentiment score of sentiment words, D represents the weight of adverbs, and S represents the sentiment score of sentences;

[0016] If the word is a sentiment word, the sentiment score P of the current sentiment word is w The sentiment score S accumulated to the current word text ; If it is an adverb, the weight D of the current adverb is a The sentiment score S accumulated to the current word text , if it is a negative word, the negation flag is set to True;

[0017] If the word does not belong to any of the above categories, the sentiment score P of the previous sentiment word is w_l Multiply by the weight D of the previous adverb a_l , and multiply it by N to get the sentiment score S of the current word text ;

[0018] If an exclamation mark is encountered, the sentiment score P of the previous sentiment word w_l , multiplied by the weight D of the previous adverb a_l , multiplied by 2, and multiplied by N to get the sentiment score S of the current word text ;

[0019] Among them, the weight of the previous adverb D a_l It is obtained by multiplying the weights of all adverbs before encountering the sentiment word; the sentiment score P of the previous sentiment word w_l It is obtained by multiplying the sentiment scores when one or more sentiment words appear consecutively.

[0020] Furthermore, the step of calculating the expression emotion score based on the emoticons in the comment data includes:

[0021] The emotional scores of different emojis were calculated according to their occurrence frequencies under positive, neutral and negative emotions, and then the three emotional scores were weighted and summed to obtain an expression emotional score, whose formula is:

[0022] S e =w 1 *S p +w 2 *S m +w 3 *S n

[0023] S p , S m , S nRepresent the occurrence frequency under positive, neutral and negative emotions respectively, w 1 、w 2 、w 3 Represents weight.

[0024] Furthermore, the final emotion score S is obtained based on the text emotion score and the expression emotion score. sen ,include:

[0025]

[0026] Among them, S norm_t Normalize the preprocessed positive text sentiment value for z-score and negative text sentiment values The collection of S norm_e The positive expression emotion value after z-score standardization preprocessing and negative expression sentiment value The collection of s and w e is the corresponding weight.

[0027] Furthermore, the text sentiment score and expression sentiment score are preprocessed by z-score standardization, and the sentiment values ​​of positive text and negative text are averaged respectively. and and variance and

[0028] Normalized preprocessed positive text sentiment value The calculation formula is as follows:

[0029]

[0030] Negative text sentiment value after normalization preprocessing The calculation formula is as follows:

[0031]

[0032] Similarly, the positive expression emotion value after standardized preprocessing is obtained and negative expression sentiment value

[0033] Furthermore, the preprocessed comment data is vectorized and compressed to obtain a text sentence vector representation, including:

[0034] The preprocessed comment data is vectorized, and each text is converted into a two-dimensional matrix. In order to unify the vector dimensions from different data, the original vector char is vectorized by the following formula sjPerform compression to unify the vector into a vector of a fixed dimension and a uniform format:

[0035] C sj =e T char sj

[0036] Among them, C sj represents the compressed vector, i.e., the text sentence vector representation; e T is a transformation matrix.

[0037] Furthermore, the text sentence vector representation C sj and the final sentiment score S sen Weighted fusion to obtain the sentence embedding representation G weighted by sentiment value sj , the formula is as follows:

[0038] G sj =S sen ·C sj .

[0039] Furthermore, the sentiment classification model includes Bert_SVM (Sentimental weighting), Roberta_SVM (Sentimental weighting) and XLNet_SVM (Sentimental weighting) models.

[0040] The present invention also provides a forum argumentative speech recognition device based on deep learning, comprising:

[0041] A data acquisition module, used to acquire and pre-process comment data, wherein the comment data includes comments by argumentative people and comments by non-argumentative people;

[0042] A first calculation module is used to calculate the text sentiment score based on the sentiment words, negative words, adverbs and exclamation marks in the preprocessed comment data;

[0043] The second calculation module is used to calculate the expression emotion score based on the emoticons in the comment data;

[0044] The third calculation module is used to obtain a final emotion score based on the text emotion score and the expression emotion score;

[0045] The vector conversion module is used to vectorize and compress the preprocessed comment data to obtain the text sentence vector representation; the text sentence vector representation and the final sentiment score are weighted and fused to obtain the sentiment value weighted sentence embedding representation;

[0046] The recognition module is used to train a sentiment classification model using sentence embedding representations weighted by sentiment values ​​and their corresponding labels; the trained sentiment classification model is used to identify whether the speech is a polemical one.

[0047] Compared with the prior art, the beneficial effects of the embodiments of the present invention are: text vector embedding is performed based on a pre-trained language model using a Transformer architecture so that it can learn richer contextual information and word relationships, and the sentiment scores of text and Emoji are calculated and weighted fused through a self-expanded sentiment dictionary so that it can effectively solve the problem of identifying argumentative remarks containing hidden contexts. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0049] Figure 1 A schematic diagram of a final sentiment score calculation method provided by an embodiment of the present invention;

[0050] Figure 2 A schematic diagram of a training data acquisition process provided by an embodiment of the present invention;

[0051] Figure 3 A flowchart of a method for identifying forum debaters' comments based on deep learning is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0053] A method for identifying forum trolls based on deep learning of the present invention, such as Figure 1-Figure 3 As shown, the specific steps are as follows:

[0054] (1) obtaining and preprocessing comment data, wherein the comment data includes comments by nitpickers and non-nitpickers;

[0055] In one embodiment, user data collection: user data collection is based on the acquisition of concentrated user samples. The experimental study of this embodiment is based on the user data of the Hupu community. First, the Hupu forum account is used to browse the hot posts of the relevant modules to filter out the users who speak and are kind to include normal user samples. The banned list is obtained in the forum related banned section and the forum search function is used to pay attention to the typical users in the banned list to obtain banned user samples (i.e., the arguing speech). According to the usage characteristics of the interface for collecting user data of the Hupu forum, a set of crawler programs is designed using Python language and Selenium automation tools to collect user data. First, the Selenium automation tool is used to imitate human operation of the browser to open the Hupu webpage login, and the WebDriver property of Selenium is used to pass the automatic verification of Hupu, and then all the user's comment data (non-arguing speech) is obtained based on the automatic drop-down module of Selenium, and finally the parsed data is saved in a file in csv format.

[0056] Preprocessing of user data: Conduct a preliminary analysis of user comment data, filter and annotate short texts, then use the Jieba tool for word segmentation. Finally, use the Harbin Institute of Technology stop word list to remove common words in the text that are irrelevant to sentiment analysis.

[0057] (2) Calculate the text sentiment score based on the sentiment words, negative words, adverbs, other words, and exclamation marks in the preprocessed comment data;

[0058] Specifically, in order to fully and comprehensively calculate the sentiment scores of these word sets, this paper obtains the sentiment scores of the text from the BosonNLP sentiment dictionary, and comprehensively considers the influence of exclamation marks and degree adverbs on the sentiment intensity, as well as the influence of negative words on the meaning of the comment to obtain the text sentiment score.

[0059] In one embodiment, the sentiment value of a short text is calculated as follows: When an exclamation mark is encountered, we calculate the sentiment score P of the previous sentiment word. w_l Multiply it by the weight D of the previous adverb a_l , and multiplied by 2 to reflect the added emotional intensity of the exclamation point.

[0060] S text _ exc =P w_l *D a_l *twenty one)

[0061] If the word is a sentiment word, the sentiment score P of the current sentiment word is w Accumulated to S text ; If it is an adverb, the weight D of the current adverb is a Accumulated to S text; If it is a negative word, set the negation flag to True.

[0062] If the word does not belong to any of the above categories (i.e., it is not an emotional word, a negative word, or an adverb), multiply the emotional score P of the previous emotional word w_l by the weight D of the previous adverb a_l , and multiply it by 1 or -1 (depending on the value of the negation flag, i.e., N), and finally accumulate it to S text , as shown in formula (2). In addition, the emotional score of the previous emotional word and the weight of the previous adverb will be reset to 0, and the negation flag will be set to false.

[0063] S text_oth = N * P w_l * D a_l (2)

[0064] If the last word is an emotional word, multiply its corresponding emotional score P w by the weight D of the previous adverb a_l , and multiply it by 1 or -1 (depending on the value of the negation flag). Finally, add the result to the text emotion S text .

[0065] S text_sen = N * P w * D a_l (3)

[0066] To further illustrate the calculation method of the text emotion score, an example is given:

[0067] Example sentence: "I really don't like this movie!"

[0068] Word segmentation result: "really don't like!"

[0069] Word type:

[0070] Adverb: really (D a is 1.3); too (D a is 1.5)

[0071] Negative word: not

[0072] Emotional word: like (emotional score is +2.1)

[0073] Calculation process: First, the degree adverb "really" is detected, and the weight is accumulated to S text as 1.3; then the negative word "not" is detected backward, the negation flag negation is set to True, and its calculated value -1 is multiplied by S text to get -1.3; when the second degree adverb "too" is detected, based on D aValue in S text Multiply and accumulate, then S text is -1.95, and the weight variable D of the preceding adverb is reset synchronously a_l is 1.95; continue to monitor backward, the sentiment score P of the sentiment word "like" w is 2.1, and the sentiment score variable P of the preceding sentiment word is reset synchronously w_l 2.1, based on P w Accumulate and get S text_word is -4.095; the exclamation mark is detected at last, based on the score of the preceding sentiment word P w_l , prepositional adverb weight variable D a_l And N gets the final comment text sentiment score S text It is -8.19.

[0074] Specifically, it can be understood that the sentiment words, negative words, adverbs and other words and exclamation marks in the comment data are assigned corresponding values, such as -1 for negative words, 1 for other types of words, and 2 for exclamation marks, and then multiplied to obtain the text sentiment score.

[0075] (3) Calculate the expression emotion score based on the emoticons in the comment data;

[0076] Specifically, we calculated the emotion scores of different emojis according to their occurrence frequencies under positive, neutral, and negative emotions, and then took the weighted sum of these three emotion scores to obtain an emotion value of the comprehensive expression. The formula is as follows:

[0077] S e =w 1 *S p +w 2 *S m +w 3 *S n (4)

[0078] (4) Based on the text sentiment score and expression sentiment score, the final sentiment score is obtained;

[0079] Specifically, the sentiment value of the text information and the sentiment value of the expression are preprocessed by z-score standardization. Taking the text as an example, the mean and variance of the sentiment values ​​of the positive text information and the negative text information are calculated respectively. The standardized positive comment sentiment value is formula (5):

[0080]

[0081] The normalized negative comment sentiment value is as follows:

[0082]

[0083] Correspondingly, the positive and negative comment sentiment values ​​after z-score standardization preprocessing are used to represent the sentiment values ​​of the expressions. and In the subsequent formula representation, we will use S to represent the standardized sentiment scores of text and expression. norm_t and S norm_e express.

[0084] Finally, the expression and text sentiment scores are weighted to obtain the final sentiment score S of the comment text. sen , the calculation formula is formula (7):

[0085]

[0086] (5) Vectorize and compress the preprocessed comment data to obtain a text sentence vector representation; weightedly fuse the text sentence vector representation and the final sentiment score to obtain a sentence embedding representation weighted by sentiment value;

[0087] Specifically, sentiment value weighted vector: The short text data is vectorized using language models BERT, RoBERTa, and XLNet. Each text is converted into a two-dimensional matrix that contains the deep semantic features of the text. Then, in order to unify the vector dimensions from different data, we apply formula (8) to compress these vectors. Here, C sj represents the compressed vector, e T is a transformation matrix, and char sj is the original vector. Through this formula, we can unify the vectors output by different models into a vector of a fixed dimension (1,768) in a unified format.

[0088] C si =e T char sj (8)

[0089] Then, the sentence embeddings are weighted and fused using the following formula to obtain the sentence embedding representation weighted by sentiment value:

[0090] G sj =S sen ·C sj (9)

[0091] (6) Use the sentence embedding representation weighted by sentiment value and its corresponding label (whether it is a polemical statement or not) to train the sentiment classification model; the trained sentiment classification model is used to identify whether it is a polemical statement or not.

[0092] Training of sentiment analysis models: Three detection models were trained using the SVM algorithm, namely SVM-Bert-Senti, SVM-Roberta-Senti, and SVM-XLNet-Senti.

[0093] Effect comparison

[0094] The experimental results of this paper are tested on a Hupu comment text information dataset collected by a crawler program, which contains more than 20,000 forum comment data. In addition, to illustrate the effectiveness of the model, it is also verified on an English dataset. In the performance evaluation of the model, the method used in this paper is to evaluate the performance of the binary confusion matrix

[0095] The method of the present invention is compared with several different types of methods for detecting trolls, especially with the insult word replacement strategy. The insult word replacement strategy is a method that aggregates all insulting content into the same symbol by replacing any insulting representation with a specific keyword (such as "fakeinsult"). The purpose of this approach is to treat all insulting speech uniformly when processing text data, so that they have the same mark in the feature representation, which makes them easier to handle in further processing and analysis. Details are as follows:

[0096] Logistic regression

[29] : Logistic regression is a classic machine learning algorithm used to solve binary classification problems. It is based on the linear regression model, but applies the Sigmoid function (also known as the Logistic function) to restrict the output between 0 and 1, thereby mapping the linear output to a probability value.

[0097] Naive Bayesian

[30] : Naive Bayesian is a probability-based classification algorithm based on Bayes’ theorem and the conditional independence assumption of features. Although its assumptions do not always hold in the real world, Naive Bayes still shows good performance in many practical problems.

[0098] SVM (Support Vector Machine)[4]: SVM is a powerful supervised learning algorithm, mainly used for binary and multi-classification problems. The core idea of ​​SVM is to find an optimal hyperplane that can effectively separate samples of different categories.

[0099] TF-IDF (Replace Insults): TF-IDF is a commonly used text feature representation method used to measure the importance of a word in a text. TF-IDF is often used in tasks such as information retrieval, text classification, and text mining. The TF-IDF method using Replace Insults can protect insulting remarks in the text while using the TF-IDF feature representation method to extract other non-insulting important features.

[0100] Naive Bayesian (Replace Insults): The Naive Bayesian method using Replace Insults can replace insulting remarks with specific keywords or symbols, thereby uniformly marking the insulting content in the feature representation, allowing the Naive Bayesian method to more focused on learning and utilizing other non-insulting features for classification.

[0101] Bert_SVM (Replace Insults): The Bert_SVM model using Replace Insults combines the Replace Insults technology with the Bert (Bidirectional Encoder Representations from Transformers) and SVM (Support Vector Machine) algorithms to handle text classification problems containing insulting remarks.

[0102] ChatGPT: ChatGPT-3.5 is a large language model based on the GPT-3.5 architecture, developed by OpenAI. As a natural language processing model, ChatGPT-3.5 can process text input and generate relevant replies. It is based on deep neural networks and uses a large amount of pre-training data to learn language patterns and semantic understanding. It continuously improves the performance of the model through iterative training and self-supervised learning.

[0103] Tables 1 and 2 compare the results of the 10 models obtained by cross-validation on two data collections, including accuracy, precision, recall, and F1-score. The higher the value, the better the recommendation effect.

[0104] Table 1 Comparison of results on the Hupu dataset

[0105]

[0106]

[0107] Table 2 Comparison of results on English datasets

[0108] Model Name A(Accuracy) P(Precision) R (Recall) F1(F1-score) Logistic regression 0.723 0.832 0.704 0.755 Naive Bayesian 0.705 0.833 0.669 0.743 SVM 0.734 0.874 0.682 0.768 TF-IDF (Replace Insults) 0.841 0.85 0.86 0.81 Naive Bayesian(Replace Insults) 0.672 0.755 0.712 0.665 Bert_SVM(Replace Insults) 0.817 0.884 0.785 0.807 Bert_SVM (Sentimental weighting) 0.886 0.798 0.779 0.793 Roberta_SVM(Sentimental weighting) 0.861 0.789 0.752 0.765 XLNet_SVM (Sentimental weighting) 0.857 0.823 0.753 0.787 ChatGPT-3.5 0.762 0.774 0.768 0.771

[0109] The present invention also provides a forum argumentative speech recognition device based on deep learning, comprising:

[0110] A data acquisition module, used to acquire and pre-process comment data, wherein the comment data includes comments by argumentative people and comments by non-argumentative people;

[0111] A first calculation module is used to calculate the text sentiment score based on the sentiment words, negative words, adverbs and exclamation marks in the preprocessed comment data;

[0112] The second calculation module is used to calculate the expression emotion score based on the emoticons in the comment data;

[0113] The third calculation module is used to obtain a final emotion score based on the text emotion score and the expression emotion score;

[0114] The vector conversion module is used to vectorize and compress the preprocessed comment data to obtain the text sentence vector representation; the text sentence vector representation and the final sentiment score are weighted and fused to obtain the sentiment value weighted sentence embedding representation;

[0115] The recognition module is used to train a sentiment classification model using sentence embedding representations weighted by sentiment values ​​and their corresponding labels; the trained sentiment classification model is used to identify whether the speech is a polemical one.

[0116] It should be noted that the device embodiment shown in this embodiment matches the content of the above method embodiment, and the content of the above method embodiment can be referred to, which will not be repeated here.

[0117] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A method for identifying forum trolls' speech based on deep learning, characterized in that: include: Obtain and preprocess comment data, wherein the comment data includes comments by argumentative people and comments by non-argumentative people; Calculate the text sentiment score based on the sentiment words, negative words, adverbs and exclamation marks in the preprocessed comment data; Calculate the expression sentiment score based on the emojis in the comment data; Based on the text sentiment score and expression sentiment score, the final sentiment score is obtained; Vectorize and compress the preprocessed comment data to obtain the text sentence vector representation; The text sentence vector representation and the final sentiment score are weighted and fused to obtain the sentiment value-weighted sentence embedding representation; The sentiment classification model is trained using the sentence embedding representation weighted by sentiment values ​​and their corresponding labels. The trained sentiment classification model is used to identify whether the speech is from a polemical person.

2. The method according to claim 1, characterized in that The pre-processing comprises: First, the short texts of the comment data are screened and annotated, then word segmentation is performed, and finally words irrelevant to sentiment analysis are removed.

3. The method according to claim 1, characterized in that: The method of calculating the text sentiment score based on sentiment words, negative words, adverbs and other words and exclamation marks in the preprocessed comment data includes: P represents the sentiment score of sentiment words, D represents the weight of adverbs, and S represents the sentiment score of sentences; If the word is a sentiment word, the sentiment score P of the current sentiment word is w The sentiment score S accumulated to the current word text ; If it is an adverb, the weight D of the current adverb is a The sentiment score S accumulated to the current word text , if it is a negative word, the negation flag is set to True; If the word does not belong to any of the above categories, the sentiment score P of the previous sentiment word is w_l Multiply by the weight D of the previous adverb a_l , get the sentiment score S of the current word text ; If an exclamation mark is encountered, the sentiment score P of the previous sentiment word w_l , multiplied by the weight D of the previous adverb a_l , multiplied by 2, and multiplied by N to get the sentiment score S of the current word text ; Among them, the weight of the previous adverb D a_l It is obtained by multiplying the weights of all adverbs before encountering the sentiment word; the sentiment score P of the previous sentiment word w_l It is obtained by multiplying the sentiment scores when one or more sentiment words appear consecutively.

4. The method according to claim 1, characterized in that The step of calculating the expression emotion score based on the emoticons in the comment data includes: The emotional scores of different emojis were calculated according to their occurrence frequencies under positive, neutral and negative emotions, and then the three emotional scores were weighted and summed to obtain an expression emotional score, whose formula is: S e =w1*S p +w2*S m +w3*S n S p , S m , S n They represent the occurrence frequencies under positive, neutral and negative emotions respectively, and w1, w2 and w3 represent the weights.

5. The method according to claim 1, characterized in that The final emotion score S is obtained based on the text emotion score and the expression emotion score. sen ,include: Among them, S norm_t Normalize the preprocessed positive text sentiment value for z-score and negative text sentiment values The collection of S norm_e The positive expression emotion value after z-score standardization preprocessing and negative expression sentiment value The collection of s and w e is the corresponding weight.

6. The method according to claim 5, characterized in that Perform z-score standardization preprocessing on the text sentiment score and expression sentiment score, and calculate the mean of the sentiment values ​​of positive text and negative text respectively and and variance and Normalized preprocessed positive text sentiment value The calculation formula is as follows: Negative text sentiment value after normalization preprocessing The calculation formula is as follows: Similarly, the positive expression emotion value after standardized preprocessing is obtained and negative expression sentiment value 7. The method according to claim 1, characterized in that The preprocessed comment data is vectorized and compressed to obtain a text sentence vector representation, including: The preprocessed comment data is vectorized, and each text is converted into a two-dimensional matrix. In order to unify the vector dimensions from different data, the original vector char is vectorized by the following formula sj Perform compression to unify the vector into a vector of a fixed dimension and a uniform format: C sj =e T character sj Among them, C sj represents the compressed vector, i.e., the text sentence vector representation; e T is a transformation matrix.

8. The method according to claim 1, characterized in that: The text sentence vector is represented by C sj and the final sentiment score S sen Weighted fusion to obtain the sentence embedding representation G weighted by sentiment value sj , the formula is as follows: G sj =S sen ·C sj 。 9. The method according to claim 1, characterized in that: The sentiment classification models include Bert_SVM, Roberta_SVM and XLNet_SVM models.

10. A device for identifying forum trolls' comments based on deep learning, characterized in that: include: A data acquisition module, used to acquire and pre-process comment data, wherein the comment data includes comments by argumentative people and comments by non-argumentative people; A first calculation module is used to calculate the text sentiment score based on the sentiment words, negative words, adverbs and exclamation marks in the preprocessed comment data; The second calculation module is used to calculate the expression emotion score based on the emoticons in the comment data; The third calculation module is used to obtain a final emotion score based on the text emotion score and the expression emotion score; The vector conversion module is used to vectorize and compress the preprocessed comment data to obtain the text sentence vector representation; The text sentence vector representation and the final sentiment score are weighted and fused to obtain the sentiment value-weighted sentence embedding representation; The recognition module is used to train a sentiment classification model using sentence embedding representations weighted by sentiment values ​​and their corresponding labels; the trained sentiment classification model is used to identify whether the speech is a polemical one.

Images