Manuscript content identification and analysis system based on natural language processing and artificial intelligence

Abstract

The invention provides a manuscript content recognition and analysis system based on natural language processing and artificial intelligence, and belongs to the field of text processing. The problem of low manuscript content recognition efficiency is solved; specifically, a manuscript obtaining module obtains a manuscript to be analyzed; the content recognition module calculates a self-attention value of each manuscript; obtaining a quasi attention value, combining the quasi attention value with the self-attention value, classifying the to-be-analyzed manuscript, and adding a content label; the content screening module is used for converting a manuscript to be analyzed into an original text vector and obtaining a marked text vector; screening false content manuscripts and bad content manuscripts in the manuscripts to be analyzed according to the original text vector and the standard text vector; the continuous updating module is used for continuously acquiring, identifying and screening the to-be-analyzed manuscripts; according to the method, the text content and the text authenticity of the manuscript are analyzed by using a self-attention mechanism, so that the recognition efficiency of the manuscript content is improved.

Description

Technical Field

[0001] The present invention discloses a manuscript content recognition and analysis system based on natural language processing and artificial intelligence, and relates to the field of text processing. Background Art

[0002] The existing manuscript content analysis system has the following shortcomings:

[0003] Data quality issues: The training effect of existing systems is highly dependent on the quality of input data. If there is noise or inaccurate labels in the data, it will lead to recognition errors or performance degradation of existing systems. If there is some bias in the training data, the existing system will also have corresponding bias.

[0004] Insufficient contextual understanding ability: Although existing AI systems can understand the grammar and semantics of text to a certain extent, they still have difficulty understanding complex contexts and implicit intentions. In the analysis of short texts, AI often lacks contextual completeness, which can cause some information to not be fully understood or misjudged.

[0005] High training costs: Advanced AI models such as deep learning require a large amount of labeled data and high computing resources for training; at the same time, the deployment and maintenance of AI systems still require expensive hardware and software resources. Summary of the invention

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a manuscript content recognition and analysis system based on natural language processing and artificial intelligence, aiming to solve the problem of low efficiency in manuscript content recognition.

[0007] In order to achieve the above object, the present invention is implemented by the following technical scheme: a manuscript content recognition and analysis system based on natural language processing and artificial intelligence includes:

[0008] Manuscript acquisition module: used to obtain the number of submission information, obtain the manuscript corresponding to each submission information, and obtain the manuscript to be analyzed;

[0009] Content identification module: used to extract all the words in the manuscript to be analyzed, and use the self-attention mechanism to calculate the dependency between words and sentences, words and paragraphs, and words and manuscripts in each manuscript to be analyzed, and obtain the self-attention value of each manuscript; obtain and calculate the self-attention value of the standard manuscript to obtain the quasi-attention value; classify the manuscripts to be analyzed according to the self-attention value and quasi-attention value of each manuscript, and add content tags to each manuscript to be analyzed;

[0010] Content screening module: used to use clustering algorithm to convert each manuscript to be analyzed into a text vector to obtain the original text vector; obtain and transform the error-corrected manuscript to obtain the marked text vector; use the cosine similarity algorithm based on the original text vector and the marked text vector to screen out the false content manuscripts and bad content manuscripts in the manuscripts to be analyzed;

[0011] Continuous update module: used to continuously obtain manuscripts to be analyzed, and to perform content identification and content screening on the newly obtained manuscripts to be analyzed.

[0012] Furthermore, the workflow of the content identification module is as follows:

[0013] Process A1: The number of content identification information is denoted as pn;

[0014] Get the number of manuscripts submitted by the first submission information, recorded as im(1); and so on, the number of manuscripts submitted by the pnth submission information, recorded as im(pn);

[0015] Calculate the sum of im(1)~im(pn), denoted as aim;

[0016] The manuscripts submitted from the 1st to the pnth submission information are recorded as mu(1,1)~mu(pn,im(pn));

[0017] Among them, mu ( 1,1 ) represents the first manuscript submitted by the first submission information; by analogy, mu(pn, im(pn)) represents the im(pn)th manuscript submitted by the pnth submission information;

[0018] Process A2: Submit the manuscript ( 1,1 ) As manuscript A; use the self-attention mechanism to calculate the dependency relationship of each word in manuscript A and get the manuscript mu ( 1,1 ) The self-attention value is denoted as aMA(1,1);

[0019] Process A3: Repeat the same process of calculating aMA(1,1), calculate the self-attention values ​​corresponding to mu(1,2)~mu(pn,im(pn)), and obtain aMA(1,2)~aMA(pn,im(pn));

[0020] Obtain scientific manuscripts, life manuscripts, shopping manuscripts, and entertainment manuscripts as standard manuscripts; repeat the same process of calculating aMA(1, 1), calculate the self-attention values ​​corresponding to scientific manuscripts, life manuscripts, shopping manuscripts, and entertainment manuscripts in turn, and obtain MA(a), MA(b), MA(c), and MA(d);

[0021] Process A4: Classify the manuscript mu(1,1) and add content tags;

[0022] Process A41: Calculate the correlation coefficient between manuscript mu(1,1) and scientific manuscript, denoted as bx(1,1-1); the calculation formula is: bx(1,1-1)=|aMA(1,1)-MA(a)| / MA(a);

[0023] The correlation coefficient between the manuscript mu(1,1) and the life manuscript is recorded as bx(1,1-2); the calculation formula is: bx(1,1-2)=|aMA(1,1)-MA(b)| / MA(b);

[0024] The correlation coefficient between the manuscript mu(1,1) and the shopping manuscript is recorded as bx(1,1-3); the calculation formula is: bx(1,1-3)=|aMA(1,1)-MA(c)| / MA(c);

[0025] The correlation coefficient between the article mu(1,1) and the entertainment article is denoted as bx(1,1-4); the calculation formula is: bx(1,1-4)=|aMA(1,1)-MA(d)| / MA(d);

[0026] Process A42: extract the minimum value among bx(1,1-1) to bx(1,1-4), denoted as bxl;

[0027] The standard manuscript type corresponding to bxl is used as the content label of the manuscript mu(1,1);

[0028] If the minimum value among bx(1,1-1)~bx(1,1-4) is not unique, the smallest correlation coefficient among bx(1,1-1)~bx(1,1-4) is recorded as bxl(1)~bxl(m);

[0029] Where bxl(1) represents the first smallest value among bx(1,1-1) to bx(1,1-4), bxl(m) represents the mth smallest value among bx(1,1-1) to bx(1,1-4), and the value range of m is 2 to 4.

[0030] The standard manuscript types corresponding to bxl(1) to bxl(m) are used as content labels of the manuscript mu(1, 1);

[0031] Content tags include: technology articles, life articles, shopping articles and entertainment articles.

[0032] Furthermore, it is characterized in that the specific process of process A2 is as follows:

[0033] Process A21: Count the number of paragraphs in manuscript A, denoted as pa;

[0034] Count the number of sentences in the first paragraph of manuscript A, recorded as se(1); and so on, the number of sentences in the pa-th paragraph is recorded as se(pa);

[0035] Process A22: Count the number of words corresponding to all sentences in manuscript A, recorded as wo(1,1)~wo(pa,se(pa));

[0036] Among them, wo(1,1) represents the number of words in the first sentence of the first paragraph; and so on, wo(pa,se(pa)) represents the number of words in the se(pa)th sentence of the path paragraph;

[0037] Extract the maximum value from wo(1,1) to wo(pa,se(pa)) and record it as wo(max), and the minimum value as wo(min); calculate the average value from wo(1,1) to wo(pa,se(pa)) and record it as wo(avg);

[0038] Process A23: Let the size of the text scanning box be tC; define relation A1:

[0039]

[0040] Determine whether the relationship A1 is satisfied and determine the calculation formula of tC;

[0041] If the relation A1 is satisfied, the calculation formula of tC is:

[0042] If the relation A1 is not satisfied, then compare the values ​​of bi(1) and bi(2) to determine the formula for calculating tC;

[0043] If bi(1)≥bi(2), then the formula for calculating tC is:

[0044] If bi(1)<bi(2), the calculation formula of tC is:

[0045] Wherein, bi(1) and bi(2) both represent transition coefficients, the calculation formula of bi(1) is: bi(1) = [(wo(avg) - wo(min)) / wo(avg)], and the calculation formula of bi(2) is: bi(2) = [(wo(max) - wo(avg)) / wo(max)];

[0046] Process A24: Process the first paragraph in manuscript A to obtain word vectors vc(1,1,1)~vc(1,se(1)1,wo(1,se(1)));

[0047] Among them, vc(1,1,1) represents the word vector of the first word in the first sentence of the first paragraph; similarly, vc(1,se(1)1,wo(1,se(1))) represents the word vector of the wo(1,se(1))th word in the se(1)th sentence of the first paragraph.

[0048] Furthermore, the subsequent process of the process A24 is as follows:

[0049] Process A25: Repeat the same process of processing the first paragraph in manuscript A, process the second to pa-th paragraphs, and obtain word vectors vc(2,1,1)~vc(pa,se(pa),wo(pa,se(pa)));

[0050] vc(2,1,1) represents the word vector of the first word in the first sentence of the second paragraph;

[0051] Similarly, vc(pa,se(pa),wo(pa,se(pa))) represents the word vector of the wo(pa,se(pa))th word in the se(pa)th sentence in the pa-th paragraph;

[0052] Process A26: Use Xavier initialization as the initialization strategy for the query-key-value weight matrix to obtain weight matrices WQ, WK, and WV;

[0053] Define calculation formula A2-1: Qi(x,y,z)=(WQ)×(vc(x,y,z));

[0054] Calculation formula A2-2: Ki(x,y,z)=(WK)×(vc(x,y,z));

[0055] Calculation formula A2-3: Vi(x,y,z)=(WV)×(vc(x,y,z));

[0056] Where × represents matrix multiplication, vc(x,y,z) represents the word vector of the zth word in the yth sentence of the xth paragraph (in manuscript A); the value range of x is 1 to pa; the value range of y is determined by x, and the value range of z is determined by x and y together;

[0057] Qi(x,y,z) represents the query vector of vc(x,y,z), Ki(x,y,z) represents the key vector of vc(x,y,z), and Vi(x,y,z) represents the value vector of vc(x,y,z);

[0058] Substituting vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) into equations A2-1 to A2-3, we can obtain the corresponding equations of vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) as follows:

[0059] Query vector: Qi(1,1,2)~Qi(pa,se(pa),wo(pa,se(pa)));

[0060] Key vector: Ki(1,1,2)~Ki(pa,se(pa),wo(pa,se(pa)));

[0061] Value vector: Vi(1,1,2)~Vi(pa,se(pa),wo(pa,se(pa)));

[0062] Process A27: Count the number of vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))), recorded as nw; nw is equal to WOB, and WOB represents the sum of wo(1,1)~wo(pa,se(pa));

[0063] Get the Unicode codes of the words corresponding to vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) and get Unz(1)~Unz(nw);

[0064] Take the word vectors vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) as quasi-vectors and get vcx(1)~vcx(nw);

[0065] Arrange the query vector, key vector and value vector corresponding to the quasi-vector in ascending order of Unz(1) to Unz(nw) to obtain: the query vector corresponding to the 1st to nwth quasi-vectors: Qix(1) to Qix(nw); key vector: Kix(1) to Kix(nw); value vector: Vix(1) to Vix(nw);

[0066] in,

[0067] Process A28: Calculate the attention score of the first quasi-vector, denoted as MA(1);

[0068] Process A29: Repeat the same process of calculating MA(1) to calculate the attention scores of the 2nd to nwth quasi-vectors to obtain MA(2)~MA(nw);

[0069] Calculate the sum of MA(1)~MA(nw), record it as aMA(1,1); take aMA(1,1) as the manuscript mu ( 1,1 ) The self-attention value.

[0070] Furthermore, it is characterized in that the specific process of the process A24 is as follows:

[0071] Process A241: process the first sentence and obtain word vectors vc(1,1,1)~vc(1,1,wo(1,1));

[0072] Set the size of the context box to tC, use the Word2Vec model to process all the words in manuscript A, and obtain the initial word vector: vca(1,1,1)~vca(pa,se(pa),wo(pa,se(pa)));

[0073] Among them, vca(1,1,1) represents the initial word vector of the first word in the first sentence of the first paragraph; and so on, vca(pa,se(pa),wo(pa,se(pa))) represents the initial word vector of the wo(pa,se(pa))th word in the se(pa)th sentence of the path paragraph;

[0074] Process A242: extract the initial word vectors of the first to wo(1,1)th words in the first sentence: vca(1,1,1)~vca(1,1,wo(1,1));

[0075] Calculate the sum of wo(1,1)~wo(1,se(1)), recorded as WOA; calculate the sum of wo(1,1)~wo(pa,se(pa)), recorded as WOB;

[0076] Process A243: cluster the initial vector vca(1,1,1) to obtain the word vector vc(1,1,1);

[0077] Process A2431: Calculate the frequency of the first word in the first to se(1)th sentences and record it as TF(1) to TF(se(1));

[0078] The word frequencies in paragraphs 1 to pa are recorded as TF(1-1)~TF(1-pa);

[0079] The frequency of all words in manuscript A is recorded as TF(1-1-e);

[0080] Calculate the inverse document frequency of the first word in all manuscripts and record it as IDF(1,1,1);

[0081] Process A2432: Calculate the average values ​​of TF(1) to TF(se(1)) and TF(1-1) to TF(1-pa) in sequence, and denote them as aTF and bTF;

[0082] Perform a linear change on vca(1,1,1) to obtain the word vector vc(1,1,1); the calculation formula of the word vector vc(1,1,1) is:

[0083]

[0084] Process A244: Repeat the same process of processing vca(1,1,1), process vca(1,1,2)~vca(1,1,wo(1,1)), and obtain word vectors vc(1,1,2)~vc(1,1,wo(1,1)).

[0085] Furthermore, the specific process of the process A28 is as follows:

[0086] Process A281: Define calculation formula A3:

[0087] Where, soc(1,j) represents the dot product of the first quasi-vector and the jth quasi-vector; Kix(j) represents the key vector of the jth quasi-vector, and the value range of j is: 2~nw;

[0088] dk represents the dimension of the weight matrix WK; × represents matrix multiplication, and T represents the transpose of the matrix;

[0089] Substitute Kix(1) to Kix(nw) into equation A3 to obtain the dot product of the first quasi-vector and the second to nwth quasi-vectors: soc(1,2) to soc(1,nw);

[0090] Process A282: Calculate the sum of soc(1,2) to soc(1,nw) and record it as so(1);

[0091] Define calculation formula A4: Among them, α(1, j) represents the weight of the first quasi-vector relative to the jth quasi-vector;

[0092] Substituting into the calculation formula A4, we can obtain the weights of the first quasi-vector relative to the second to nw quasi-vectors: α(1,2)~α(1,nw);

[0093] Process A282: Calculate the attention score MA(1) of the first quasi-vector according to soc(1,2)~soc(1,nw) and α(1,2)~α(1,nw). The calculation formula of MA(1) is:

[0094]

[0095] Furthermore, the workflow of the content screening module is as follows:

[0096] Process B1: Use K-means clustering algorithm to convert manuscripts mu(1, 1)~mu(pn, im(pn)) into text vectors to obtain text vectors ve(1, 1)~ve(pn, im(pn));

[0097] Among them, the vector ve(1,1) represents the text vector corresponding to the manuscript mu(1,1); the vector ve(pn,im(pn)) represents the text vector corresponding to the manuscript mu(pn,im(pn));

[0098] Process B2: obtaining a manuscript of false content as error correction manuscript A, and obtaining a manuscript of bad content as error correction manuscript B;

[0099] The number of statistically corrected manuscripts A is recorded as wa, and the number of statistically corrected manuscripts B is recorded as wb;

[0100] Using the K-means clustering algorithm, the 1st to wath error correction manuscripts A are converted into text vectors to obtain text vectors wve(1, 1)~wve(1, wa);

[0101] Convert the 1st to wbth error correction manuscripts A into text vectors, and obtain text vectors wve(2,1)~wve(2,wb);

[0102] Process B3: According to the vectors ve(1,1)~ve(pn,im(pn)) and the vectors wve(1,1)~wve(1,wa), the manuscripts with false contents in the manuscripts mu(1,1)~mu(pn,im(pn)) are eliminated;

[0103] Process B4: Repeat the same process of judging whether the manuscripts mu(1, 1)~mu(pn, im(pn)) are manuscripts with false content, and eliminate the manuscripts with bad content from the manuscripts mu(1, 1)~mu(pn, im(pn)) based on ve(1, 1)~ve(pn, im(pn)) and vectors wve(2, 1)~wve(2, wb).

[0104] Furthermore, the specific process of process B3 is as follows:

[0105] Process B31: Determine whether the manuscript mu(1, 1) is a false content manuscript;

[0106] Calculate the dot product of manuscript mu(1,1) with respect to the error correction manuscript A, denoted as jd; the calculation formula of jd is: Wherein, × represents matrix multiplication, T represents matrix transpose; wve(1, j) represents the text vector of the jth error correction manuscript A, and the value range of j is: 1~wa;

[0107] Process B32: Calculate the norm of the vector ve(1,1), denoted as fa; the calculation formula of fa is:

[0108] fa=|ve( 1,1 )× ( ve( 1,1 ) )T |;

[0109] Calculate the average norm of wve(1,1)~wve(1,wa), recorded as fb; the calculation formula of fb is:

[0110]

[0111] Process B33: Calculate the cosine similarity of vector ve(1,1) with respect to vectors wve(1,1) to wve(1,wa), denoted as csi; the calculation formula of csi is: csi=jd / (fa×fb);

[0112] Determine whether (1-csi) is less than ε; where ε represents the determination coefficient;

[0113] If (1-csi) is greater than or equal to ε, it means that the manuscript mu(1,1) is not a false content manuscript;

[0114] If (1-csi) is less than ε, it means that the manuscript mu(1,1) is a false content manuscript;

[0115] Process B34: Repeat the same process of judging whether the manuscript mu(1,1) is a false content manuscript, and remove the false content manuscripts from the manuscripts mu(1,2)~mu(pn,im(pn)) based on ve(1,2)~ve(pn,im(pn)) and vectors wve(1,1)~wve(1,wa);

[0116] Among them, the vector ve(1, 2) represents the text vector corresponding to the manuscript mu(1, 2).

[0117] Compared with the prior art, the present invention has the following beneficial effects:

[0118] Efficiency and automation: The present invention can process a large amount of document data, reducing the time and cost of manual operations; the present invention conducts deep learning and mining of the text of the manuscript, allowing the system to identify the potential connections and relationships between words, between words and paragraphs, and between words and manuscripts from massive data, thereby reducing subjective bias in manual identification.

[0119] Support large-scale data processing: The present invention can extract information from massive data, adapt to the needs of the big data era, and automatically identify and process ultra-large-scale texts.

[0120] Improve the quality of text analysis: The system of the present invention can process and analyze a large amount of manuscript content in a very short time. Especially in scenarios with high real-time requirements (such as social media monitoring), the present invention can identify and analyze content in real time, and accurately classify, screen and label manuscript content, providing a reference for media workers. BRIEF DESCRIPTION OF THE DRAWINGS

[0121] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0122] Figure 1 It is a schematic diagram of the system of the present invention;

[0123] Figure 2 This is a schematic diagram of data processing of the present invention;

[0124] Figure 3 Schematic diagram of self-attention value calculation of the present invention. DETAILED DESCRIPTION

[0125] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0126] See also Figure 1 and Figure 2 ,The manuscript content recognition and analysis system based on natural language processing and artificial intelligence includes: manuscript acquisition module, content recognition module, content screening module and continuous update module;

[0127] Manuscript acquisition module: used to obtain the number of submission information, obtain the manuscript corresponding to each submission information, and obtain the manuscript to be analyzed;

[0128] It should be noted that the "submission information" in the present invention means: the identity of the contributor or the contributing unit who submits the manuscript to the present invention (the manuscript content recognition and analysis system based on natural language processing and artificial intelligence);

[0129] Content identification module: used to extract all the words in the manuscript to be analyzed, and use the self-attention mechanism to calculate the dependency between words and sentences, words and paragraphs, and words and manuscripts in each manuscript to be analyzed, and obtain the self-attention value of each manuscript; obtain and calculate the self-attention value of the standard manuscript to obtain the quasi-attention value; classify the manuscripts to be analyzed according to the self-attention value and quasi-attention value of each manuscript, and add content tags to each manuscript to be analyzed;

[0130] Process A: Please refer to Figure 3 ,The workflow of the content identification module is as follows:

[0131] Process A1: The number of content identification information is denoted as pn;

[0132] Get the number of manuscripts submitted by the first submission information, recorded as im(1); the number of manuscripts submitted by the second submission information, recorded as im(2); and so on, the number of manuscripts submitted by the pnth submission information, recorded as im(pn);

[0133] Calculate the sum of im(1)~im(pn), denoted as aim;

[0134] The manuscripts submitted from the 1st to the pnth submission information are recorded as mu(1,1), mu(1,2)~mu(1,im(1))~mu(pn,im(pn));

[0135] Among them, mu ( 1,1 ) represents the first manuscript submitted by the first submission; mu(1, 2) represents the second manuscript submitted by the first submission; and so on, mu(1, im(1)) represents the im(1)th manuscript submitted by the first submission; mu(pn, im(pn)) represents the im(pn)th manuscript submitted by the pnth submission;

[0136] Process A2: Submit the manuscript ( 1,1 ) As manuscript A; use the self-attention mechanism to calculate the dependency relationship of each word in manuscript A and get the manuscript mu ( 1,1 ) The self-attention value is denoted as aMA(1,1);

[0137] Process A21: Count the number of paragraphs in manuscript A, denoted as pa;

[0138] Count the number of sentences in the first paragraph of manuscript A, recorded as se(1); the number of sentences in the second paragraph, recorded as se(2); and so on, the number of sentences in the pa-th paragraph, recorded as se(pa);

[0139] Process A22: Count the number of words corresponding to all sentences in manuscript A, recorded as wo(1,1), wo(1,2)~wo(1,se(1))~wo(pa,se(pa));

[0140] Among them, wo(1,1) represents the number of words in the first sentence of the first paragraph (in manuscript A); wo(1,2) represents the number of words in the second sentence of the first paragraph (in manuscript A); and so on, wo(1,se(1)) represents the number of words in the se(1)th sentence of the first paragraph (in manuscript A); wo(pa,se(pa)) represents the number of words in the se(pa)th sentence of the path paragraph (in manuscript A);

[0141] Extract the maximum value from wo(1,1) to wo(pa,se(pa)) and record it as wo(max), and the minimum value as wo(min); calculate the average value from wo(1,1) to wo(pa,se(pa)) and record it as wo(avg);

[0142] Process A23: Let the size of the text scanning box be tC; define relation A1:

[0143]

[0144] Determine whether the relationship A1 is satisfied and determine the calculation formula of tC;

[0145] If the relation A1 is satisfied, the calculation formula of tC is:

[0146] If the relation A1 is not satisfied, then compare the values ​​of bi(1) and bi(2) to determine the formula for calculating tC;

[0147] If bi(1)≥bi(2), then the formula for calculating tC is:

[0148] If bi(1)<bi(2), the calculation formula of tC is:

[0149] Wherein, bi(1) and bi(2) both represent transition coefficients, the calculation formula of bi(1) is: bi(1) = [(wo(avg) - wo(min)) / wo(avg)], and the calculation formula of bi(2) is: bi(2) = [(wo(max) - wo(avg)) / wo(max)];

[0150] Process A24: Process the first paragraph in manuscript A to obtain word vectors vc(1,1,1), vc(1,1,2)~vc(1,1,wo(1,1))~vc(1,se(1)1,wo(1,se(1)));

[0151] Among them, vc(1,1,1) represents the word vector of the first word in the first sentence of the first paragraph (in manuscript A); vc(1,1,2) represents the word vector of the second word in the first sentence of the first paragraph (in manuscript A);

[0152] Similarly, vc(1,1,wo(1,1)) represents the word vector of the wo(1,1)th word in the first sentence of the first paragraph (in manuscript A);

[0153] vc(1,se(1)1,wo(1,se(1))) represents the word vector of the wo(1,se(1))th word in the se(1)th sentence in the first paragraph (in manuscript A);

[0154] Process A241: process the first sentence (of the first paragraph in manuscript A) and obtain word vectors vc(1,1,1), vc(1,1,2)~vc(1,1,wo(1,1));

[0155] Set the size of the context box to tC, use the Word2Vec model to process all the words in manuscript A, and obtain the initial word vectors: vca(1,1,1), vca(1,1,2)~vca(1,1,wo(1,1))~vca(pa,se(pa),wo(pa,se(pa)));

[0156] Among them, vca(1,1,1) represents the initial word vector of the first word in the first sentence of the first paragraph (in manuscript A); vca(1,1,2) represents the initial word vector of the second word in the first sentence of the first paragraph (in manuscript A);

[0157] Similarly, vca(1,1,wo(1,1)) represents the initial word vector of the wo(1,1)th word in the first sentence of the first paragraph (in manuscript A);

[0158] vca(pa,se(pa),wo(pa,se(pa))) represents the initial word vector of the wo(pa,se(pa))th word in the se(pa)th sentence in the pa-th paragraph (in manuscript A);

[0159] Process A242: extract the initial word vectors of the first to wo(1,1)th words in the first sentence (of the first paragraph in manuscript A): vca(1,1,1)~vca(1,1,wo(1,1));

[0160] Calculate the sum of wo(1,1)~wo(1,se(1)), recorded as WOA; calculate the sum of wo(1,1)~wo(pa,se(pa)), recorded as WOB;

[0161] Process A243: cluster the initial vector vca(1,1,1) to obtain the word vector vc(1,1,1);

[0162] Process A2431: Calculate the word frequency of the first word (in the first sentence of the first paragraph of manuscript A) in the first to se(1)th sentences (in the first paragraph of manuscript A), recorded as TF(1) to TF(se(1));

[0163] The frequency of the first word in the first sentence of the first paragraph of manuscript A in the first to pa paragraphs of manuscript A is recorded as TF(1-1) to TF(1-pa);

[0164] The frequency of the first word in the first sentence of the first paragraph of manuscript A is recorded as TF(1-1-e);

[0165] Calculate the inverse document frequency of the first word (in the first sentence of the first paragraph in manuscript A) in all manuscripts (i.e., the first to aimth manuscripts) and record it as IDF(1,1,1);

[0166] Process A2432: Calculate the average values ​​of TF(1) to TF(se(1)) and TF(1-1) to TF(1-pa) in sequence, and denote them as aTF and bTF;

[0167] Perform a linear change on vca(1,1,1) to obtain the word vector vc(1,1,1); the calculation formula of the word vector vc(1,1,1) is:

[0168]

[0169] Process A244: Repeat the same process of processing vca(1,1,1) (i.e., process A243), process vca(1,1,2)~vca(1,1,wo(1,1)), and obtain word vectors vc(1,1,2)~vc(1,1,wo(1,1));

[0170] Process A25: Repeat the same process of processing the first paragraph in manuscript A (i.e., process A24), process the second to pa-th paragraphs (in manuscript A), and obtain word vectors vc(2,1,1)~vc(pa,se(pa),wo(pa,se(pa)));

[0171] vc(2,1,1) represents the word vector of the first word in the first sentence of the second paragraph (in manuscript A);

[0172] By analogy, vc(pa,se(pa),wo(pa,se(pa))) represents the word vector of the wo(pa,se(pa))th word in the se(pa)th sentence in the pa-th paragraph (in manuscript A);

[0173] Process A26: Use Xavier initialization as the initialization strategy for the query-key-value weight matrix to obtain weight matrices WQ, WK, and WV;

[0174] Define calculation formula A2-1: Qi(x,y,z)=(WQ)×(vc(x,y,z));

[0175] Calculation formula A2-2: Ki(x,y,z)=(WK)×(vc(x,y,z));

[0176] Calculation formula A2-3: Vi(x,y,z)=(WV)×(vc(x,y,z));

[0177] Wherein, × represents matrix multiplication, vc(x,y,z) represents the word vector of the zth word in the yth sentence of the xth paragraph (in manuscript A); the value range of x is 1 to pa; the value range of y is determined by x (for example, if x is 1, the value range of y is 1 to se(1); if x is 2, the value range of y is 1 to se(2); by analogy, if x is pa, the value range of y is 1 to se(pa);), and the value range of z is determined by both x and y; for example, if x is 1 and y is 1, the value range of z is 1 to wo(1,1); if x is 1 and y is 2, the value range of z is 1 to wo(1,2); by analogy, if x is pa and y is se(pa), the value range of z is 1 to wo(pa,se(pa));

[0178] Qi(x,y,z) represents the query vector of vc(x,y,z), Ki(x,y,z) represents the key vector of vc(x,y,z), and Vi(x,y,z) represents the value vector of vc(x,y,z);

[0179] Substituting vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) into equations A2-1 to A2-3, we can obtain the corresponding equations of vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) as follows:

[0180] Query vector: Qi(1,1,2)~Qi(pa,se(pa),wo(pa,se(pa)));

[0181] Key vector: Ki(1,1,2)~Ki(pa,se(pa),wo(pa,se(pa)));

[0182] Value vector: Vi(1,1,2)~Vi(pa,se(pa),wo(pa,se(pa)));

[0183] Process A27: Count the number of vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))), recorded as nw; nw is equal to WOB, and WOB represents the sum of wo(1,1)~wo(pa,se(pa));

[0184] Get the Unicode codes of the words corresponding to vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) and get Unz(1)~Unz(nw);

[0185] Take the word vectors vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) as quasi-vectors and get vcx(1)~vcx(nw);

[0186] Arrange the query vector, key vector and value vector corresponding to the quasi-vector in ascending order of Unz(1) to Unz(nw) to obtain: the query vector corresponding to the 1st to nwth quasi-vectors: Qix(1) to Qix(nw); key vector: Kix(1) to Kix(nw); value vector: Vix(1) to Vix(nw);

[0187] in,

[0188] Process A28: Calculate the attention score of the first quasi-vector, denoted as MA(1);

[0189] Process A281: Define calculation formula A3:

[0190] Where, soc(1,j) represents the dot product of the first quasi-vector and the jth quasi-vector; Kix(j) represents the key vector of the jth quasi-vector, and the value range of j is: 2~nw;

[0191] dk represents the dimension of the weight matrix WK; × represents matrix multiplication, and T represents the transpose of the matrix;

[0192] Substitute Kix(1) to Kix(nw) into equation A3 to obtain the dot product of the first quasi-vector and the second to nwth quasi-vectors: soc(1,2) to soc(1,nw);

[0193] Process A282: Calculate the sum of soc(1,2) to soc(1,nw) and record it as so(1);

[0194] Define calculation formula A4: Among them, α(1, j) represents the weight of the first quasi-vector relative to the jth quasi-vector;

[0195] Substituting into the calculation formula A4, we can obtain the weights of the first quasi-vector relative to the second to nw quasi-vectors: α(1,2)~α(1,nw);

[0196] Process A282: Calculate the attention score MA(1) of the first quasi-vector according to soc(1,2)~soc(1,nw) and α(1,2)~α(1,nw). The calculation formula of MA(1) is:

[0197]

[0198] Process A29: Repeat the same process of calculating MA(1) to calculate the attention scores of the 2nd to nwth quasi-vectors to obtain MA(2)~MA(nw);

[0199] Calculate the sum of MA(1)~MA(nw), record it as aMA(1,1); take aMA(1,1) as the manuscript mu ( 1,1 ) The self-attention value of

[0200] Process A3: Repeat the same process of calculating aMA(1,1), calculate the self-attention values ​​corresponding to mu(1,2)~mu(pn,im(pn)), and obtain aMA(1,2)~aMA(pn,im(pn));

[0201] (Online) obtain (existing) scientific manuscripts, life manuscripts, shopping manuscripts and entertainment manuscripts as standard manuscripts; repeat the same process of calculating aMA(1,1), calculate the self-attention values ​​corresponding to scientific manuscripts, life manuscripts, shopping manuscripts and entertainment manuscripts in turn, and obtain MA(a), MA(b), MA(c) and MA(d);

[0202] Process A4: Classify the manuscript mu(1,1) and add content tags;

[0203] Process A41: Calculate the correlation coefficient between manuscript mu(1,1) and scientific manuscript, denoted as bx(1,1-1); the calculation formula is: bx(1,1-1)=|aMA(1,1)-MA(a)| / MA(a);

[0204] The correlation coefficient between the manuscript mu(1,1) and the life manuscript is recorded as bx(1,1-2); the calculation formula is: bx(1,1-2)=|aMA(1,1)-MA(b)| / MA(b);

[0205] The correlation coefficient between the manuscript mu(1,1) and the shopping manuscript is recorded as bx(1,1-3); the calculation formula is: bx(1,1-3)=|aMA(1,1)-MA(c)| / MA(c);

[0206] The correlation coefficient between the article mu(1,1) and the entertainment article is denoted as bx(1,1-4); the calculation formula is: bx(1,1-4)=|aMA(1,1)-MA(d)| / MA(d);

[0207] Process A42: extract the minimum value among bx(1,1-1) to bx(1,1-4), denoted as bxl;

[0208] The standard manuscript type corresponding to bxl is used as the content label of the manuscript mu(1,1);

[0209] If the minimum value among bx(1,1-1)~bx(1,1-4) is not unique, the smallest correlation coefficient among bx(1,1-1)~bx(1,1-4) is recorded as bxl(1)~bxl(m);

[0210] Where bxl(1) represents the first smallest value among bx(1,1-1) to bx(1,1-4), bxl(m) represents the mth smallest value among bx(1,1-1) to bx(1,1-4), and the value range of m is 2 to 4.

[0211] The standard manuscript types corresponding to bxl(1) to bxl(m) are used as content labels of the manuscript mu(1,1);

[0212] Content tags include: science and technology articles, life articles, shopping articles and entertainment articles;

[0213] It should be noted that the present invention only lists four specific types of labels, namely "technology, life, shopping, and entertainment". Media workers or relevant technical personnel can define or design more labels on their own when using the present invention to improve the benefits of the present invention to different user groups.

[0214] Content screening module: used to use clustering algorithm to convert each manuscript to be analyzed into a text vector to obtain the original text vector; obtain and transform the error-corrected manuscript to obtain the marked text vector; use the cosine similarity algorithm based on the original text vector and the marked text vector to screen out the false content manuscripts and bad content manuscripts in the manuscripts to be analyzed;

[0215] Process B: The workflow of the content screening module is as follows:

[0216] Process B1: Use K-means clustering algorithm to convert manuscripts mu(1,1)~mu(pn,im(pn)) into text vectors, and obtain text vectors ve(1,1)~ve(pn,im(pn)); (original text vector);

[0217] Among them, the vector ve(1,1) represents the text vector corresponding to the manuscript mu(1,1); the vector ve(pn,im(pn)) represents the text vector corresponding to the manuscript mu(pn,im(pn));

[0218] Process B2: (Online or from the Internet regulatory department) obtain the manuscript of false content as the error correction manuscript A, and obtain the manuscript of bad content as the error correction manuscript B;

[0219] The number of statistically corrected manuscripts A is recorded as wa, and the number of statistically corrected manuscripts B is recorded as wb;

[0220] Using the K-means clustering algorithm, the 1st to wath error correction manuscripts A are converted into text vectors, and the text vectors wve(1,1)~wve(1,wa) are obtained; (labeled text vectors);

[0221] Convert the 1st to wbth error correction manuscripts A into text vectors, and obtain text vectors wve(2,1)~wve(2,wb); (mark text vector);

[0222] Process B3: According to the vectors ve(1,1)~ve(pn,im(pn)) and the vectors wve(1,1)~wve(1,wa), the manuscripts with false contents in the manuscripts mu(1,1)~mu(pn,im(pn)) are eliminated;

[0223] Process B31: Determine whether the manuscript mu(1, 1) is a false content manuscript;

[0224] Calculate the dot product of manuscript mu(1,1) with respect to the error correction manuscript A, denoted as jd; the calculation formula of jd is: Wherein, × represents matrix multiplication, T represents matrix transpose; wve(1, j) represents the text vector of the jth error correction manuscript A, and the value range of j is: 1~wa;

[0225] Process B32: Calculate the norm of the vector ve(1,1), denoted as fa; the calculation formula of fa is:

[0226] fa=|ve( 1,1 )× ( ve( 1,1 ) )T |;

[0227] Calculate the average norm of wve(1,1)~wve(1,wa), recorded as fb; the calculation formula of fb is:

[0228]

[0229] Process B33: Calculate the cosine similarity of vector ve(1,1) with respect to vectors wve(1,1) to wve(1,wa), denoted as csi; the calculation formula of csi is: csi=jd / (fa×fb);

[0230] Determine whether (1-csi) is less than ε; where ε represents the determination coefficient; the value of ε is generally 0.2; users or relevant technical personnel can adjust the value of ε according to actual needs;

[0231] If (1-csi) is greater than or equal to ε, it means that the manuscript mu(1,1) is not a false content manuscript;

[0232] If (1-csi) is less than ε, it means that the manuscript mu(1,1) is a false content manuscript;

[0233] Process B34: Repeat the same process (i.e., process B31 to process B33) of determining whether the manuscript mu(1,1) is a false content manuscript, and remove the false content manuscripts from the manuscripts mu(1,2) to mu(pn,im(pn)) based on ve(1,2) to ve(pn,im(pn)) and vectors wve(1,1) to wve(1,wa);

[0234] Among them, the vector ve(1, 2) represents the text vector corresponding to the manuscript mu(1, 2);

[0235] Process B4: Repeat the same process (i.e., process B3) of judging whether the manuscripts mu(1, 1) to mu(pn, im(pn)) are manuscripts with false content, and eliminate the manuscripts with bad content from the manuscripts mu(1, 1) to mu(pn, im(pn)) based on ve(1, 1) to ve(pn, im(pn)) and vectors wve(2, 1) to wve(2, wb).

[0236] Continuous update module: used to continuously obtain manuscripts to be analyzed, and to perform content identification and content screening on the newly obtained manuscripts to be analyzed.

[0237] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technical personnel in this field according to actual conditions. For example, if there are weight coefficients and proportional coefficients, their set sizes are to quantify each parameter to obtain a specific value, which is convenient for subsequent comparison. Regarding the size of the weight coefficient and the proportional coefficient, it is sufficient as long as it does not affect the proportional relationship between the parameter and the quantized value.

[0238] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A manuscript content recognition and analysis system based on natural language processing and artificial intelligence, characterized by: The system comprises: Manuscript acquisition module: used to obtain the number of submission information, obtain the manuscript corresponding to each submission information, and obtain the manuscript to be analyzed; Content identification module: used to extract all the words in the manuscript to be analyzed, and use the self-attention mechanism to calculate the dependency between words and sentences, words and paragraphs, and words and manuscripts in each manuscript to be analyzed, and obtain the self-attention value of each manuscript; obtain and calculate the self-attention value of the standard manuscript to obtain the quasi-attention value; classify the manuscripts to be analyzed according to the self-attention value and quasi-attention value of each manuscript, and add content tags to each manuscript to be analyzed; Content screening module: used to use clustering algorithm to convert each manuscript to be analyzed into a text vector to obtain the original text vector; obtain and transform the error-corrected manuscript to obtain the marked text vector; use the cosine similarity algorithm based on the original text vector and the marked text vector to screen out the false content manuscripts and bad content manuscripts in the manuscripts to be analyzed; Continuous update module: used to continuously obtain manuscripts to be analyzed, and to perform content identification and content screening on the newly obtained manuscripts to be analyzed.

2. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 1 is characterized in that: The workflow of the content identification module is as follows: Process A1: The number of content identification information is denoted as pn; Get the number of manuscripts submitted by the first submission information, recorded as im(1); and so on, the number of manuscripts submitted by the pnth submission information, recorded as im(pn); Calculate the sum of im(1)~im(pn), denoted as aim; The manuscripts submitted from the 1st to the pnth submission information are recorded as mu(1,1)~mu(pn,im(pn)); Among them, mu ( 1,1 ) represents the first manuscript submitted by the first submission information; by analogy, mu(pn, im(pn)) represents the im(pn)th manuscript submitted by the pnth submission information; Process A2: Submit the manuscript ( 1,1 ) As manuscript A; use the self-attention mechanism to calculate the dependency relationship of each word in manuscript A and get the manuscript mu ( 1,1 ) The self-attention value is denoted as aMA(1,1); Process A3: Repeat the same process of calculating aMA(1,1), calculate the self-attention values ​​corresponding to mu(1,2)~mu(pn,im(pn)), and obtain aMA(1,2)~aMA(pn,im(pn)); Obtain scientific manuscripts, life manuscripts, shopping manuscripts, and entertainment manuscripts as standard manuscripts; repeat the same process of calculating aMA(1, 1), calculate the self-attention values ​​corresponding to scientific manuscripts, life manuscripts, shopping manuscripts, and entertainment manuscripts in turn, and obtain MA(a), MA(b), MA(c), and MA(d); Process A4: Classify the manuscript mu(1,1) and add content tags; Process A41: Calculate the correlation coefficient between manuscript mu(1,1) and scientific manuscript, denoted as bx(1,1-1); the calculation formula is: bx(1,1-1)=|aMA(1,1)-MA(a)| / MA(a); The correlation coefficient between the manuscript mu(1,1) and the life manuscript is recorded as bx(1,1-2); the calculation formula is: bx(1,1-2)=|aMA(1,1)-MA(b)| / MA(b); The correlation coefficient between the manuscript mu(1,1) and the shopping manuscript is recorded as bx(1,1-3); the calculation formula is: bx(1,1-3)=|aMA(1,1)-MA(c)| / MA(c); The correlation coefficient between the article mu(1,1) and the entertainment article is denoted as bx(1,1-4); the calculation formula is: bx(1,1-4)=|aMA(1,1)-MA(d)| / MA(d); Process A42: extract the minimum value among bx(1,1-1) to bx(1,1-4), denoted as bxl; The standard manuscript type corresponding to bxl is used as the content label of the manuscript mu(1,1); If the minimum value among bx(1,1-1)~bx(1,1-4) is not unique, the smallest correlation coefficient among bx(1,1-1)~bx(1,1-4) is recorded as bxl(1)~bxl(m); Where bxl(1) represents the first smallest value among bx(1,1-1) to bx(1,1-4), bxl(m) represents the mth smallest value among bx(1,1-1) to bx(1,1-4), and the value range of m is 2 to 4. The standard manuscript types corresponding to bxl(1) to bxl(m) are used as content labels of the manuscript mu(1, 1); Content tags include: technology articles, life articles, shopping articles and entertainment articles.

3. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 2 is characterized in that: The specific process of process A2 is as follows: Process A21: Count the number of paragraphs in manuscript A, denoted as pa; Count the number of sentences in the first paragraph of manuscript A, recorded as se(1); and so on, the number of sentences in the pa-th paragraph is recorded as se(pa); Process A22: Count the number of words corresponding to all sentences in manuscript A, recorded as wo(1,1)~wo(pa,se(pa)); Among them, wo(1,1) represents the number of words in the first sentence of the first paragraph; and so on, wo(pa,se(pa)) represents the number of words in the se(pa)th sentence of the path paragraph; Extract the maximum value from wo(1,1) to wo(pa,se(pa)) and record it as wo(max), and the minimum value as wo(min); calculate the average value from wo(1,1) to wo(pa,se(pa)) and record it as wo(avg); Process A23: Let the size of the text scanning box be tC; define relation A1: Determine whether the relationship A1 is satisfied and determine the calculation formula of tC; If the relation A1 is satisfied, the calculation formula of tC is: If the relation A1 is not satisfied, then compare the values ​​of bi(1) and bi(2) to determine the formula for calculating tC; If bi(1)≥bi(2), then the formula for calculating tC is: If bi(1)<bi(2), the calculation formula of tC is: Wherein, bi(1) and bi(2) both represent transition coefficients, the calculation formula of bi(1) is: bi(1) = [(wo(avg) - wo(min)) / wo(avg)], and the calculation formula of bi(2) is: bi(2) = [(wo(max) - wo(avg)) / wo(max)]; Process A24: Process the first paragraph in manuscript A to obtain word vectors vc(1,1,1)~vc(1,se(1)1,wo(1,se(1))); Among them, vc(1,1,1) represents the word vector of the first word in the first sentence of the first paragraph; similarly, vc(1,se(1)1,wo(1,se(1))) represents the word vector of the wo(1,se(1))th word in the se(1)th sentence of the first paragraph.

4. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 3 is characterized in that: The subsequent process of the process A24 is as follows: Process A25: Repeat the same process of processing the first paragraph in manuscript A, process the second to pa-th paragraphs, and obtain word vectors vc(2,1,1)~vc(pa,se(pa),wo(pa,se(pa))); vc(2,1,1) represents the word vector of the first word in the first sentence of the second paragraph; Similarly, vc(pa,se(pa),wo(pa,se(pa))) represents the word vector of the wo(pa,se(pa))th word in the se(pa)th sentence in the pa-th paragraph; Process A26: Use Xavier initialization as the initialization strategy for the query-key-value weight matrix to obtain weight matrices WQ, WK, and WV; Define calculation formula A2-1: Qi(x,y,z)=(WQ)×(vc(x,y,z)); Calculation formula A2-2: Ki(x,y,z)=(WK)×(vc(x,y,z)); Calculation formula A2-3: Vi(x,y,z)=(WV)×(vc(x,y,z)); Where × represents matrix multiplication, vc(x,y,z) represents the word vector of the zth word in the yth sentence of the xth paragraph; the value range of x is 1 to pa; the value range of y is determined by x, and the value range of z is determined by x and y together; Qi(x,y,z) represents the query vector of vc(x,y,z), Ki(x,y,z) represents the key vector of vc(x,y,z), and Vi(x,y,z) represents the value vector of vc(x,y,z); Substituting vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) into equations A2-1 to A2-3, we can obtain the corresponding equations of vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) as follows: Query vector: Qi(1,1,2)~Qi(pa,se(pa),wo(pa,se(pa))); Key vector: Ki(1,1,2)~Ki(pa,se(pa),wo(pa,se(pa))); Value vector: Vi(1,1,2)~Vi(pa,se(pa),wo(pa,se(pa))); Process A27: Count the number of vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))), recorded as nw; Get the Unicode codes of the words corresponding to vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) and get Unz(1)~Unz(nw); Take the word vectors vc(1,1,2)~vc(pa,se(pa),wo(pa,se(pa))) as quasi-vectors and get vcx(1)~vcx(nw); Arrange the query vector, key vector and value vector corresponding to the quasi-vector in ascending order of Unz(1) to Unz(nw) to obtain: the query vector corresponding to the 1st to nwth quasi-vectors: Qix(1) to Qix(nw); key vector: Kix(1) to Kix(nw); value vector: Vix(1) to Vix(nw); Process A28: Calculate the attention score of the first quasi-vector, denoted as MA(1); Process A29: Repeat the same process of calculating MA(1) to calculate the attention scores of the 2nd to nwth quasi-vectors to obtain MA(2)~MA(nw); Calculate the sum of MA(1)~MA(nw), record it as aMA(1,1); take aMA(1,1) as the manuscript mu ( 1,1 ) The self-attention value.

5. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 3 is characterized in that: The specific process of process A24 is as follows: Process A241: process the first sentence and obtain word vectors vc(1,1,1)~vc(1,1,wo(1,1)); Set the size of the context box to tC, use the Word2Vec model to process all the words in manuscript A, and obtain the initial word vector: vca(1,1,1)~vca(pa,se(pa),wo(pa,se(pa))); Among them, vca(1,1,1) represents the initial word vector of the first word in the first sentence of the first paragraph; and so on, vca(pa,se(pa),wo(pa,se(pa))) represents the initial word vector of the wo(pa,se(pa))th word in the se(pa)th sentence of the path paragraph; Process A242: extract the initial word vectors of the first to wo(1,1)th words in the first sentence: vca(1,1,1)~vca(1,1,wo(1,1)); Calculate the sum of wo(1,1)~wo(1,se(1)), recorded as WOA; calculate the sum of wo(1,1)~wo(pa,se(pa)), recorded as WOB; Process A243: Cluster the initial vector vca(1,1,1) to obtain the word vector vc(1,1,1); Process A244: Repeat the same process of processing vca(1,1,1), process vca(1,1,2)~vca(1,1,wo(1,1)), and obtain word vectors vc(1,1,2)~vc(1,1,wo(1,1)).

6. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 4 is characterized in that: The specific process of process A28 is as follows: Process A281: Define calculation formula A3: Where, soc(1,j) represents the dot product of the first quasi-vector and the jth quasi-vector; Kix(j) represents the key vector of the jth quasi-vector, and the value range of j is: 2~nw; dk represents the dimension of the weight matrix WK; × represents matrix multiplication, and T represents the transpose of the matrix; Substitute Kix(1) to Kix(nw) into equation A3 to obtain the dot product of the first quasi-vector and the second to nwth quasi-vectors: soc(1,2) to soc(1,nw); Process A282: Calculate the sum of soc(1,2) to soc(1,nw) and record it as so(1); Define calculation formula A4: Among them, α(1, j) represents the weight of the first quasi-vector relative to the jth quasi-vector; Substituting into the calculation formula A4, we can obtain the weights of the first quasi-vector relative to the second to nw quasi-vectors: α(1,2)~α(1,nw); Process A282: Calculate the attention score MA(1) of the first quasi-vector according to soc(1,2)~soc(1,nw) and α(1,2)~α(1,nw). The calculation formula of MA(1) is:

7. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 1 is characterized in that: The workflow of the content screening module is as follows: Process B1: Use K-means clustering algorithm to convert manuscripts mu(1, 1)~mu(pn, im(pn)) into text vectors to obtain text vectors ve(1, 1)~ve(pn, im(pn)); Among them, the vector ve(1,1) represents the text vector corresponding to the manuscript mu(1,1); the vector ve(pn,im(pn)) represents the text vector corresponding to the manuscript mu(pn,im(pn)); Process B2: obtaining a manuscript of false content as error correction manuscript A, and obtaining a manuscript of bad content as error correction manuscript B; The number of statistically corrected manuscripts A is recorded as wa, and the number of statistically corrected manuscripts B is recorded as wb; Using the K-means clustering algorithm, the 1st to wath error correction manuscripts A are converted into text vectors to obtain text vectors wve(1, 1)~wve(1, wa); Convert the 1st to wbth error correction manuscripts A into text vectors, and obtain text vectors wve(2,1)~wve(2,wb); Process B3: According to the vectors ve(1,1)~ve(pn,im(pn)) and the vectors wve(1,1)~wve(1,wa), the manuscripts with false contents in the manuscripts mu(1,1)~mu(pn,im(pn)) are eliminated; Process B4: Repeat the same process of judging whether the manuscripts mu(1, 1)~mu(pn, im(pn)) are manuscripts with false content, and eliminate the manuscripts with bad content from the manuscripts mu(1, 1)~mu(pn, im(pn)) based on ve(1, 1)~ve(pn, im(pn)) and vectors wve(2, 1)~wve(2, wb).

8. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 7 is characterized in that: The specific process of process B3 is as follows: Process B31: Determine whether the manuscript mu(1, 1) is a false content manuscript; Calculate the dot product of manuscript mu(1,1) with respect to the error correction manuscript A, denoted as jd; the calculation formula of jd is Wherein, × represents matrix multiplication, T represents matrix transpose; wve(1, j) represents the text vector of the jth error correction manuscript A, and the value range of j is: 1~wa; Process B32: Calculate the norm of the vector ve(1,1), denoted as fa; the calculation formula of fa is: fa=|and( 1,1 )× ( And( 1,1 ) )T |; Calculate the average norm of wve(1,1)~wve(1,wa), denoted as fb; the calculation formula of fb is: Process B33: Calculate the cosine similarity of vector ve(1,1) with respect to vectors wve(1,1) to wve(1,wa), denoted as csi; the calculation formula of csi is: csi=jd / (fa×fb); Determine whether (1-csi) is less than ε; where ε represents the determination coefficient; If (1-csi) is greater than or equal to ε, it means that the manuscript mu(1,1) is not a false content manuscript; If (1-csi) is less than ε, it means that the manuscript mu(1,1) is a false content manuscript; Process B34: Repeat the same process of judging whether the manuscript mu(1,1) is a false content manuscript, and remove the false content manuscripts from the manuscripts mu(1,2)~mu(pn,im(pn)) based on ve(1,2)~ve(pn,im(pn)) and vectors wve(1,1)~wve(1,wa); Among them, the vector ve(1, 2) represents the text vector corresponding to the manuscript mu(1, 2).

9. The manuscript content recognition and analysis system based on natural language processing and artificial intelligence according to claim 5 is characterized in that: The details of the process A243 are as follows: Process A2431: Calculate the frequency of the first word in the first to se(1)th sentences and record it as TF(1) to TF(se(1)); The word frequencies in paragraphs 1 to pa are recorded as TF(1-1)~TF(1-pa); The frequency of all words in manuscript A is recorded as TF(1-1-e); Calculate the inverse document frequency of the first word in all manuscripts and record it as IDF(1,1,1); Process A2432: Calculate the average values ​​of TF(1) to TF(se(1)) and TF(1-1) to TF(1-pa) in sequence, and denote them as aTF and bTF; Perform a linear change on vca(1,1,1) to obtain the word vector vc(1,1,1); the calculation formula of the word vector vc(1,1,1) is:

Images