A multi-level multi-dimensional feature fusion false review analysis method

Abstract

This invention provides a multi-level, multi-dimensional feature fusion method for analyzing fake reviews. This method comprises a word-level feature processing module, a sentence-level feature processing module, a review-level feature processing module, and a global feature processing module. This innovative method proposes a multi-level, multi-dimensional feature fusion model for analyzing fake reviews. By connecting various granularities of reviews through a multi-level network structure, it fuses multi-dimensional features and integrates semantics and auxiliary information at each layer. This effectively solves the problems of existing methods that fail to deeply study the internal granularity relationships and feature processing differences within reviews, have single-dimensional feature extraction, and struggle to comprehensively capture review information, making it difficult to distinguish between genuine and fake reviews in complex cases. This improves the effectiveness of fake review identification.

Description

Technical Field

[0001] This invention is a method for analyzing fake reviews by falsifying multi-level and multi-dimensional features, specifically a technology for detecting and identifying fake reviews on e-commerce platforms. Background Technology

[0002] The development of the internet and e-commerce has made online reviews a key factor in consumer decisions; however, a large number of fake reviews mislead consumers and damage platform credibility. Existing methods do not delve into the internal granular relationships and feature processing differences within reviews, making it difficult to comprehensively capture review information and distinguish between genuine and fake reviews in complex cases. To address this, this invention proposes a multi-level, multi-dimensional feature fusion method. This method uses a multi-level network structure to connect information at various granularities within reviews, fusing multi-dimensional features and integrating semantic and auxiliary information at each layer to improve the effectiveness of fake review identification. Summary of the Invention

[0003] The present invention provides a multi-level, multi-dimensional feature fusion method for analyzing fake comments, comprising: a word-level feature processing module, a sentence-level feature processing module, a comment-level feature processing module, and a global feature processing module. The overall processing flow is as follows: Figure 1 As shown.

[0004] The processing flow of the word-level feature processing module of this invention is as follows: First, read the comment text to be processed and perform sentence segmentation and word segmentation operations on the text; Second, embed vectorized representations of each word in each sentence; Third, extract and fuse word-level language features of each word in each sentence; Fourth, fuse the vectorized representations of each word with the word-level language features to obtain the text feature representation of each word; Fifth, determine whether the current word is the last word of the sentence. If so, input the result into the sentence-level feature processing module; otherwise, return to the second step.

[0005] The processing flow of the sentence-level feature processing module of the present invention is as follows: First, obtain the text feature representation of each word in each sentence output by the word-level feature processing module; Second, fuse the text feature representation of each word in each sentence, and then calculate the semantic feature representation of the sentence according to formulas (1)-(3); Third, perform vectorization representation on each sentence to obtain the original vector representation of each sentence, and retain the original features of the sentence; Fourth, extract and fuse the sentence-level language features of each sentence; Fifth, fuse the sentence-level language features, the original vector representation of the sentence, and the semantic feature representation of the sentence to obtain a precise fusion. The sentence-level features; sixth, based on the precisely fused sentence-level features and the three different weight matrices of sentence investigation, labeling, and start, the investigation vector, labeling vector, and result vector of each sentence are calculated by formulas (4)-(6); seventh, based on the calculated investigation vector, labeling vector, and result vector of each sentence, the attention optimization representation of each sentence is calculated by formula (7), and the attention optimization representation is used as the final text feature representation of each sentence; eighth, determine whether the current sentence is the last sentence of the comment. If so, input the result into the comment-level feature processing module; otherwise, return to execute the first step.

[0006] The processing flow of the comment-level feature processing module of the present invention is as follows: First, obtain the text feature representation of all sentences in the comment output by the sentence-level feature processing module; Second, fuse the text feature representation of all sentences in the comment, and then calculate the text semantic feature representation of the comment according to formulas (8)-(10); Third, vectorize the comment text to obtain the original vector representation of the comment, and retain the original features of the comment; Fourth, extract and fuse the comment-level language features of the comment; Fifth, fuse the comment-level language features of the comment text, the original vector representation of the comment, and the semantic feature representation of the comment to obtain the precisely fused comment-level features; Sixth, according to the precisely fused comment-level features and the three different weight matrices of comment investigation, labeling, and start, calculate the investigation vector, labeling vector, and result vector of the comment respectively through formulas (11)-(13); Seventh, according to the calculated investigation vector, labeling vector, and result vector of the comment, calculate the multi-head attention optimization representation of the comment through formula (14), and use the multi-head attention optimization representation as the final text feature representation of the comment; Eighth, input the result into the global feature processing module.

[0007] The processing flow of the global feature processing module of the present invention is as follows: First, obtain the text feature representation of the comment output by the comment-level feature processing module; Second, extract and fuse the user action features of the comment; Third, extract and fuse the item features of the comment; Fourth, fuse the text feature representation of the comment, the user action features of the comment, and the item features of the comment to obtain the precisely fused global feature of the comment; Fifth, calculate the final global depth feature representation of the comment using formulas (15)-(17) based on the precisely fused global feature of the comment; Sixth, perform feature-classification space mapping on the global depth feature representation of the comment to obtain the category prediction result of the comment.

[0008] The calculation formula for the fake review analysis method of this invention is defined as follows:

[0009] (1) Formula for calculating the contextual semantic features of a single comment sentence

[0010] (1)

[0011] (2)

[0012] (3)

[0013] Where || represents the concatenation operation.

[0014] (2) Formulas for calculating the single-sentence query vector, the single-sentence tag vector, and the single-sentence result vector.

[0015] (4)

[0016]

[0017]

[0018] (3) Calculation formula for optimizing the attention of a single comment sentence

[0019] (7)

[0020] (4) Formula for calculating the semantic features of the full text context of the comment

[0021] (8)

[0022] (9)

[0023] (10)

[0024] (5) Formulas for calculating the full-text search vector, the full-text tag vector, and the full-text result vector of the comments.

[0025] (11) Comment full text tag vector; comment full text tag weight matrix; comment full text comprehensive representation vector

[0026] (13)

[0027] (6) Calculation formula for optimizing the full text attention of comments

[0028] (14)

[0029] (7) Global deep feature calculation formula

[0030] (15)

[0031] (16)

[0032] (17)

[0033] Specific processing steps of the present invention

[0034] The steps of the word-level feature processing module, sentence-level feature processing module, comment-level feature processing module, and global feature processing module of the analysis method of this invention are as follows.

[0035] like Figure 2 As shown, the processing flow of the word-level feature processing module is as follows:

[0036] Start at page 201;

[0037] P202 Enter comment text;

[0038] Page 203 performs sentence segmentation and word segmentation on the comment text;

[0039] P204 embeds a vectorized representation of each word;

[0040] P205 Extracts and merges word-level language features for each word;

[0041] P206 integrates the vectorized representation of each word with word-level language features to obtain the text feature representation of each word;

[0042] P207 checks if the current word is the last word of the sentence. If it is, proceed to P208 sequentially; otherwise, return to P204.

[0043] P208 Input the results into the sentence-level feature processing module.

[0044] Page 209, End.

[0045] like Figure 3As shown, the processing flow of the sentence-level feature processing module is as follows:

[0046] Start at page 301;

[0047] P302 Reads the text feature representation of each word in each sentence output by the word-level feature processing module;

[0048] P303 integrates the text feature representation of each word in each sentence and calculates the semantic feature representation of each sentence according to formulas (1)-(3);

[0049] P304 Vectorize each sentence to obtain the original vector representation of the sentence;

[0050] P305 Extracts and fuses sentence-level language features for each sentence;

[0051] P306 integrates the sentence-level linguistic features, the original vector representation of the sentence, and the semantic feature representation of the sentence to obtain accurately integrated sentence-level features;

[0052] P307 Based on the precisely fused sentence-level features and formulas (4)-(6), calculate the single-sentence query vector, tag vector, and result vector of the comment;

[0053] P308 Calculate the attention optimization representation of each sentence based on the single sentence query vector, tag vector and result vector, and formula (7), as the final text feature representation of each sentence;

[0054] P309 checks if the current sentence is the last sentence of the comment. If it is, proceed to P310 sequentially; otherwise, return to P302.

[0055] P310 Input the results into the comment-level feature processing module.

[0056] End of page 311.

[0057] like Figure 4 As shown, the processing flow of the comment-level feature processing module is as follows:

[0058] Start at page 401;

[0059] P402 Reads the text feature representations of all sentences output by the sentence-level feature processing module;

[0060] P403 The text feature representations of all sentences in the comment are integrated, and the text semantic feature representation of the entire comment is calculated according to formulas (8)-(10);

[0061] P404 Vectorizes the comment text to obtain the original vector representation of the comment;

[0062] P405 Extract and merge comment-level linguistic features of comments;

[0063] P406 integrates the comment-level linguistic features of the comment text, the original vector representation of the comment, and the semantic feature representation of the comment to obtain a precisely integrated comment-level feature;

[0064] P407 Based on the precisely fused comment-level features and formulas (11)-(13), calculate the full-text search vector, tag vector, and result vector of the comment respectively;

[0065] P408 Calculate the multi-head attention optimization representation of the comment based on the full-text search vector, tag vector and result vector of the comment, as well as formula (14), and use it as the final text feature representation of the comment;

[0066] P409 Input the results into the global feature processing module.

[0067] End of P410.

[0068] like Figure 5 As shown, the processing flow of the global feature processing module is as follows:

[0069] Start at page 501;

[0070] P502 Reads the text feature representation of the comment output by the comment-level feature processing module;

[0071] P503 Extracts and merges user action features from comments;

[0072] P504 Extracts and merges item features from comments;

[0073] P505 integrates the textual features of the comment, the user action features of the comment, and the item features of the comment to obtain a precisely integrated global feature of the comment;

[0074] P506 The final global depth feature representation of the comments is calculated based on the precisely fused global features of the comments and formulas (15)-(17);

[0075] P507 performs a feature-classification space mapping on the global deep feature representation of the comment to obtain the category prediction result of the comment.

[0076] Page 508 is the end. Attached Figure Description

[0078] Figure 1 This is the overall processing flowchart of the present invention;

[0079] Figure 2 This is a flowchart of the word-level feature processing module of the present invention;

[0080] Figure 3 This is a flowchart of the sentence-level feature processing module of the present invention;

[0081] Figure 4 This is a flowchart of the comment-level feature processing module of the present invention;

[0082] Figure 5 This is a flowchart of the global feature processing module of the present invention; Detailed Implementation

[0084] The specific implementation of the multi-level, multi-dimensional feature fusion method for analyzing fake reviews according to the present invention is as follows.

[0085] Step 1: Execute the "Word-level Feature Processing Module"

[0086] The English product review texts analyzed in this embodiment of the invention are from the Yelp review dataset.

[0087] (1) Enter comment text:

[0088] User_id: 932; Product_id: 1; Review_date: 2014 / 5 / 20; Rating: 2

[0089] Review_text: “The photos of this desk lamp are very misleading. The 'warm white' light is actually a harsh, cold blue and the adjustable arm is sostiff it's difficult to position. The build quality feels cheap and flimsy.Would not recommend for serious use.”

[0090] (2) The comment text to be analyzed was segmented into sentences and words, and the results are as follows:

[0091] The first sentence: ['The', 'photos', 'of', 'this', 'desk', 'lamp', 'are', 'very', 'misleading', '.']

[0092] The second sentence: ['The', " ' ", 'warm', 'white', " ' ", 'light', 'is', 'actually', 'a', 'harsh', ',']

[0093] The third sentence: ['cold', 'blue', 'and', 'the', 'adjustable', 'arm', 'is', 'so', 'stiff', 'it', " 's ", 'difficult', 'to', 'position', '.']

[0094] The fourth sentence: ['The', 'build', 'quality', 'feels', 'cheap', 'and', 'flimsy', '.']

[0095] The fifth sentence: ['Would', 'not', 'recommend', 'for', 'serious', 'use', '.']

[0096] (3) Embedded vector representation of each word in each sentence, the result is as follows:

[0097] Vectorized representation of the first word in the first sentence:

[0098] [0.8807, -1.2684, 0.1905, ..., 0.2199, 0.1187, -0.1230] ...

[0100] Vectorized representation of the last word in the first sentence:

[0101] [1.0305, -1.4199, 0.2066, ..., 0.0157, -0.0346, -0.1913] .... .... ....

[0103] Vectorized representation of the first word in the fifth sentence:

[0104] [0.4963, -0.2068, 0.8723, ..., -0.2827, 0.0982, 0.4054] ...

[0106] Vectorized representation of the last word in the fifth sentence:

[0107] [0.6180, 2.1412, -0.4407, ..., 0.5559, 0.8620, -0.7033]

[0108] (4) For each word in each sentence, extract and fuse the word-level language features. The results are as follows:

[0109] Word-level linguistic features of the first word in the first sentence:

[0110] [-1.1994, 0.1255, -1.2051, ..., 0.5381, 0.0000, 0.0000] ...

[0112] Word-level linguistic features of the last word in the first sentence:

[0113] [0.5014, -0.9264, -0.7075, ..., 1.7264, -0.4146, -0.3360] .... .... ....

[0115] Word-level linguistic features of the first word in the fifth sentence:

[0116] [1.2500, 1.0621, 0.5011, ..., -0.7304, -1.5594, 0.9015] ...

[0118] Word-level linguistic features of the last word in the fifth sentence:

[0119] [-0.9063, 2.2242, -1.6432, ..., 0.1555, 0.2218, 0.0000]

[0120] (5) The vectorized representation of each word is fused with the word-level language features to obtain the text feature representation of each word. The results are as follows:

[0121] Text feature representation of the first word in the first sentence:

[0122] [0.8807, -1.2684, 0.1905, ..., 0.5381, 0.0000, 0.0000] ...

[0124] Text feature representation of the last word in the first sentence:

[0125] [1.0305, -1.4199, 0.2066, ..., 1.7264, -0.4146, -0.3360] .... .... ....

[0127] Text feature representation of the first word in the fifth sentence:

[0128] [0.4963, -0.2068, 0.8723, ..., -0.7304, -1.5594, 0.9015] ...

[0130] Text feature representation of the last word in the fifth sentence:

[0131] [0.6180, 2.1412, -0.4407, ..., 0.1555, 0.2218, 0.0000]

[0132] Step 2: Execute the "Sentence-Level Feature Processing Module"

[0133] (1) The text feature representations of each word in each sentence are fused together, and the semantic feature representation of each sentence is calculated according to formulas (1)-(3). The results are as follows:

[0134] Semantic feature representation of the first sentence in the comment text:

[0135] [[-0.0357, 0.7852, 0.0141, ..., 0.0288, -0.0256, -0.6297],

[0136] [-0.1102, 1.1658, 0.0221, ..., 0.0546, -0.0460, -0.9392], ...

[0138] [-0.3867, 0.5556, 0.0153, ..., -0.2076, -0.0660, -0.9514],

[0139] [-0.1632, 0.1707, 0.0396, ..., -0.1083, 0.0856, -0.4047]] .... .... ....

[0141] Semantic feature representation of the fifth sentence in the comment text:

[0142] [[-0.0375, 0.7591, -0.0376, ..., 0.0625, -0.0629, -0.6136],

[0143] [-0.1379, 1.1843, 0.0212, ..., 0.0605, -0.1260, -0.9478], ...

[0145] [-0.4630, 0.6164, -0.0142, ..., -0.2197, -0.0604, -0.7475],

[0146] [-0.1557, 0.2226, -0.0006, ..., -0.1288, 0.0284, -0.2546]]

[0147] (2) Vectorize each sentence to obtain the original vector representation of the sentence. The results are as follows:

[0148] The original vector representation of the first sentence in the comment text:

[0149] [ 0.0108, -0.0333, 0.0298, ..., -0.0194, 0.0443, 0.1190,

[0150] -0.0213, 0.0333, 0.0331, ..., 0.1178, -0.0689, -0.2887, ...

[0152] -0.1312, -0.0632, 0.1210, ..., 0.2795, -0.0180, -0.0061,

[0153] -0.1205, -0.1302, 0.2042, ..., -0.1822, -0.0032, 0.0283 .... .... ....

[0155] The original vector representation of the fifth sentence in the comment text:

[0156] [ 0.0289, -0.0796, 0.0128, ..., -0.0176, 0.0820, 0.1054,

[0157] 0.0029, 0.0556, 0.0222, ..., 0.1403, -0.0722, -0.2796, ...

[0159] -0.1000, -0.0827, 0.0822, ..., 0.2593, -0.0409, -0.0218,

[0160] -0.0931, -0.1237, 0.1922, ..., -0.2052, 0.0131, -0.0235

[0161] (3) For each sentence, extract and fuse the sentence-level language features of each sentence, and the results are as follows:

[0162] Sentence-level linguistic features of the first sentence in the comment text:

[0163] [ 0.1019, 0.1000, 0.2268, ..., 0.0963, 0.1425, 0.0000,

[0164] 0.0064, 0.0000, 0.0000, ..., 0.0000, 0.2143, 0.1540, ...

[0166] 0.0000, 0.1959, 0.0286, ..., 0.0455, 0.0000, 0.0000,

[0167] 0.0000, 0.1354, 0.0948, ..., 0.0762, 0.0000, 0.0008 .... .... ....

[0169] Sentence-level linguistic features of the fifth sentence in the comment text:

[0170] [ 0.0906, 0.0673, 0.1950, ..., 0.0741, 0.1361, 0.0000,

[0171] 0.0000, 0.0000, 0.0000, ..., 0.0141, 0.1764, 0.1376,

[0172] ... ...,

[0173] 0.0000, 0.1760, 0.0489, ..., 0.0313, 0.0000, 0.0000,

[0174] 0.0000, 0.1035, 0.0740, ..., 0.0777, 0.0000, 0.0000

[0175] (4) The sentence-level linguistic features, the original vector representation of the sentence, and the semantic feature representation of the sentence are fused together to obtain the precisely fused sentence-level features, as shown below:

[0176] The sentence-level features of the first sentence in the comment text are precisely fused:

[0177] [ [0.0060, 0.2386, 0.0000, ..., 0.0458, 0.0815, 0.0000],

[0178] [-0.0357, 0.7852, 0.0141, ..., 0.0288, -0.0256, -0.6297], ...

[0180] [-0.3867, 0.5556, 0.0153, ..., -0.2076, -0.0660, -0.9514],

[0181] [-0.1632, 0.1707, 0.0396, ..., -0.1083, 0.0856, -0.4047] .... .... ....

[0183] The fifth sentence in the comment text has precisely fused sentence-level features:

[0184] [ [0.0000, 0.2248, 0.0000, ..., 0.0517, 0.0643, 0.0000],

[0185] [-0.0375, 0.7591, -0.0376, ..., 0.0625, -0.0629, -0.6136], ...

[0187] [-0.4630, 0.6164, -0.0142, ..., -0.2197, -0.0604, -0.7475],

[0188] [-0.1557, 0.2226, -0.0006, ..., -0.1288, 0.0284, -0.2546]

[0189] (5) Based on the precisely fused sentence-level features and formulas (4)-(6), the single-sentence query vector, tag vector, and result vector of the comment are calculated respectively, and the results are as follows:

[0190] The single-sentence query vector Q of the first sentence in the comment text:

[0191] [[0.0631, 0.0281, -0.0142, ..., 0.0313, -0.0249, -0.0487],

[0192] [0.0529, 0.0887, -0.1280, ..., -0.0299, -0.0305, -0.0169], ...

[0194] [0.0631, 0.0281, -0.0142, ..., 0.0313, -0.0249, -0.0487],

[0195] [0.0529, 0.0887, -0.1280, ..., -0.0299, -0.0305, -0.0169]] .... .... ....

[0197] The single-sentence query vector Q of the fifth sentence in the comment text:

[0198] [[0.0603, 0.0279, -0.0156, ..., 0.0304, -0.0229, -0.0473],

[0199] [0.0091, -0.1252, -0.0863, ..., 0.2200, -0.0721, -0.0056],

[0200] .... ...,

[0201] [0.0371, -0.1464, -0.0614, ..., 0.3685, -0.0055, 0.0081],

[0202] [0.0438, -0.0565, -0.0277, ..., 0.1747, -0.0132, -0.0210]]

[0203] The single-sentence tag vector K of the first sentence in the comment text:

[0204] [[-0.0198, -0.0184, -0.0296, ..., -0.0095, 0.0213, 0.0606],

[0205] [0.0072, -0.1172, -0.0332, ..., 0.0381, 0.2143, -0.1313], ...

[0207] [0.0409, -0.0688, -0.0881, ..., -0.0054, 0.2596, -0.3052],

[0208] [[-0.0032, -0.0256, -0.0490, ..., 0.0236, 0.0737, -0.1298]] .... .... ....

[0210] The single-sentence tag vector K of the fifth sentence in the comment text:

[0211] [[-0.0171, -0.0171, -0.0296, ..., -0.0097, 0.0233, 0.0582],

[0212] [-0.0119, -0.0962, -0.0340, ..., 0.0368, 0.1898, -0.1370], ...

[0214] [0.0584, -0.0749, -0.0277, ..., -0.0217, 0.2043, -0.2585],

[0215] [-0.0022, -0.0483, -0.0478, ..., 0.0046, 0.0636, -0.0951]]

[0216] The single-sentence result vector V of the first sentence in the comment text:

[0217] [[-0.0198, -0.0184, -0.0296, ..., -0.0095, 0.0213, 0.0606],

[0218] [0.0072, -0.1172, -0.0332, ..., 0.0381, 0.2143, -0.1313], ...

[0220] [0.0409, -0.0688, -0.0881, ..., -0.0054, 0.2596, -0.3052],

[0221] [[-0.0032, -0.0256, -0.0490, ..., 0.0236, 0.0737, -0.1298]] .... .... ....

[0223] The single-sentence result vector V of the fifth sentence in the comment text:

[0224] [[0.0260, -0.0364, 0.0350, ..., 0.0631, 0.1170, 0.0484],

[0225] [-0.6956, 0.1636, -0.1864, ..., -0.2424, -0.0935, -0.2873], ...

[0227] [0.0584, -0.0749, -0.0277, ..., -0.0217, 0.2043, -0.2585],

[0228] [-0.0022, -0.0483, -0.0478, ..., 0.0046, 0.0636, -0.0951]]

[0229] (6) Based on the single-sentence query vector, tag vector, and result vector of the comment, and formula (7), calculate the attention optimization representation of each sentence as the final text feature representation of each sentence. The results are as follows:

[0230] The final text feature representation of the first sentence in the comment text:

[0231] [ [-2.5679, 0.9205, -0.6359, ..., -0.9102, -0.2003, -0.5071],

[0232] [-1.8461, 1.4216, -0.3880, ..., -0.6194, -0.2296, -1.1377], ...

[0234] [-2.0017, 1.0061, -0.3195, ..., -0.7920, -0.2269, -1.3479],

[0235] [-2.3000, 0.7346, -0.3907, ..., -0.9038, -0.0818, -0.9547] .... .... ....

[0237] The final text feature representation of the fifth sentence in the comment text:

[0238] [ [-2.5578, 0.9308, -0.5153, ..., -0.8410, -0.1572, -0.6844],

[0239] [-1.7444, 1.4761, -0.3518, ..., -0.4956, -0.2359, -1.3191], ...

[0241] [-2.1419, 1.1656, -0.2748, ..., -0.8133, -0.1973, -1.3668],

[0242] [-2.3239, 0.8465, -0.3507, ..., -0.9395, -0.1153, -0.9493]

[0243] Step 3: Execute the "Comment-Level Feature Processing Module"

[0244] (1) The text feature representations of all sentences in the comment are integrated, and the text semantic feature representation of the entire comment is calculated according to formulas (8)-(10). The results are as follows:

[0245] Textual semantic feature representation of comment text:

[0246] [[0.0968, 0.2219, 0.0473, ..., -0.3247, -0.5143, 0.5280],

[0247] [0.1155, 0.3446, 0.0788, ..., -0.3377, -0.5419, 0.5690], ...

[0249] [0.1139, 0.4251, 0.1251, ..., -0.2100, -0.3321, 0.3116],

[0250] [0.1045, 0.3738, 0.1201, ..., -0.1230, -0.1842, 0.1654]]

[0251] (2) Vectorize the comment text to obtain the original vector representation of the comment, as shown below:

[0252] The original vector representation of the comment text:

[0253] [0.7265, -0.8633, 0.2445, ..., -0.2878, 0.1370, 0.0086,

[0254] 0.1461, 1.0909, -0.7824, ..., -0.7062, 0.9894, 0.0131, ...

[0256] -1.5264, 1.6894, -0.2155, ..., -0.4001, -0.2632, -1.4588,

[0257] -2.0364, 1.3017, -0.2727, ..., -0.7722, -0.3198, -1.5388]

[0258] (3) For each comment, the comment-level linguistic features of the comment are extracted and fused, and the results are as follows:

[0259] Comment-level linguistic features of comment texts:

[0260] [ 0.0674, 0.0300, 0.0171, ..., 0.0000, 0.0000, 0.0012,

[0261] 0.1161, 0.2241, 0.0000, ..., 0.0000, 0.0000, 0.0857, ...

[0263] 0.0001, 0.1209, 0.0658, ..., 0.0000, 0.0998, 0.0000,

[0264] 0.1470, 0.0700, 0.0000, ..., 0.0000, 0.0000, 0.1207

[0265] (4) The comment-level linguistic features, the original vector representation of the comment, and the semantic feature representation of the comment are fused together to obtain the precisely fused comment-level features, as shown below:

[0266] Comment-level features that are precisely integrated into the comment text:

[0267] [ [0.0000, 0.1209, 0.0258, ..., -0.2123, 0.0000, 0.2744],

[0268] [0.0968, 0.2219, 0.0473, ..., -0.3247, -0.5143, 0.5280], ...

[0270] [0.1139, 0.4251, 0.1251, ..., -0.2100, -0.3321, 0.3116],

[0271] [0.1045, 0.3738, 0.1201, ..., -0.1230, -0.1842, 0.1654]

[0272] (5) Based on the precise fusion of comment-level features and formulas (11)-(13), the comment full-text search vector, tag vector, and result vector are calculated respectively, and the results are as follows:

[0273] Comment text full text search vector Q:

[0274] [[0.1218, -0.1576, 0.0075, ..., -0.4553, -0.1445, -0.2185],

[0275] [-0.0697, 0.0174, 0.1016, ..., -0.0008, -0.1061, -0.0128], ...

[0277] -0.1051, 0.0115, 0.0537, ..., 0.0846, -0.0149, -0.0177],

[0278] [-0.0948, 0.0102, 0.0307, ​​..., 0.0901, 0.0203, -0.0094]]

[0279] Comment text full-text tag vector K:

[0280] [[0.1218, -0.1576, 0.0075, ..., -0.4553, -0.1445, -0.2185],

[0281] [-0.0697, 0.0174, 0.1016, ..., -0.0008, -0.1061, -0.0128], ...

[0283] [-0.1051, 0.0115, 0.0537, ..., 0.0846, -0.0149, -0.0177],

[0284] [-0.0948, 0.0102, 0.0307, ​​..., 0.0901, 0.0203, -0.0094][

[0285] The full-text result vector V of the comment text:

[0286] [[0.2085, 0.0918, -0.2990, ..., 0.0325, 0.2067, -0.1930],

[0287] [-0.1730, 0.2722, 0.0723, ..., -0.2384, -0.1944, 0.4357], ...

[0289] [-0.1682, 0.3987, 0.0782, ..., -0.1659, -0.3190, 0.4500],

[0290] [-0.1352, 0.3365, 0.0801, ..., -0.0888, -0.2609, 0.3380]]

[0291] (6) Based on the full-text search vector, tag vector, and result vector of the comment, and formula (14), the multi-head attention optimization representation of the comment is calculated as the final text feature representation of the comment. The results are as follows:

[0292] The final text feature representation of the comment text:

[0293] [ [-0.8898, 0.3031, -0.4138, ..., -1.0052, -1.1645, 0.4785],

[0294] [-0.2626, 1.3950, 0.2031, ..., -1.2139, -1.7740, 2.2024], ...

[0296] [-0.2272, 1.7058, 0.3305, ..., -0.9512, -1.3577, 1.6827],

[0297] [-0.2613, 1.7273, 0.3454, ..., -0.8431, -1.1557, 1.5093]

[0298] Step 5: Execute the "Global Feature Processing Module"

[0299] (1) For the entire comment, the user action features of the comment are extracted and fused, and the results are as follows:

[0300] User behavior characteristics of this comment:

[0301] [0.0000, 0.0000, 0.1862, ..., 0.0236, 0.0000, 0.3481,

[0302] 0.1295, 0.1052, 0.0000, ..., 0.0000, 0.1532, 0.1032, ...

[0304] 0.0000, 0.0000, 0.0000, ..., 0.0460, 0.0000, 0.1973,

[0305] 0.0000, 0.0306, 0.0082, ..., 0.0462, 0.0000, 0.0310]

[0306] (2) For the entire review, the item features of the review are extracted and merged, and the results are as follows:

[0307] The item characteristics described in this review:

[0308] [0.0000, 0.0000, 0.1213, ..., 0.0000, 0.0126, 0.0358,

[0309] 0.0000, 0.0000, 0.0000, ..., 0.0472, 0.0000, 0.0097, ...

[0311] 0.1976, 0.0355, 0.0000, ..., 0.1669, 0.0188, 0.0592,

[0312] 0.1027, 0.0920, 0.0384, ..., 0.0524, 0.0000, 0.0000]

[0313] (3) The textual features of the comment, the user action features of the comment, and the item features of the comment are fused together to obtain the precise fused global features of the comment, as shown below:

[0314] The commentary precisely integrates the following global features:

[0315] [[0.0000, 0.0679, 0.0000, ..., 0.0000, 0.6348, 0.0273],

[0316] [0.0000, 0.0000, 0.5804, ..., 1.2315, 0.9870, 0.4450], ...

[0318] [0.0000, 0.0233, 2.5719, ..., 4.4879, 0.1000, 0.3300],

[0319] [0.0000, 0.0269, 2.4809, ..., 4.3742, 0.0896, 0.2204]]

[0320] (4) Based on the precisely fused global features of the comments and formulas (15)-(17), the final global deep feature representation of the comments is calculated, and the results are as follows:

[0321] The global deep feature representation of this comment:

[0322] [0.0000, 0.0000, 5.3684, ..., 3.1300, 4.7317, 5.0302,

[0323] 0.4973, 0.0000, 0.0000, ..., 0.0049, 0.0014, 4.4661, ...

[0325] 9.7679, 0.0000, 0.0000, ..., 1.0026, 0.0000, 0.0002,

[0326] 0.0000, 0.0000, 0.0147, ..., 0.0000, 6.4773, 0.0000]

[0327] (5) Map the global deep feature representation of the comments to the feature-classification space to obtain the category prediction results of the comments, as follows:

[0328] Map the global deep feature representation to a binary classification logical score:

[0329] The comment's true / false logic score: [-0.0956, 0.1456]

[0330] Convert logical scores to categorical probabilities:

[0331] The probability of this comment being a true / false prediction: [0.44, 0.56]

[0332] Predict based on probability output:

[0333] Is this comment predicted to be fake? Yes.

Claims

1. A method for analyzing fake reviews based on multi-level and multi-dimensional feature fusion, characterized by: ... The system consists of a word-level feature processing module, a sentence-level feature processing module, a comment-level feature processing module, and a global feature processing module. Its fake comment analysis method includes the following processing flow: (1) The processing flow of the word-level feature processing module is as follows: First, read the comment text to be processed and perform sentence segmentation and word segmentation on the text; Second, embed vector representation of each word in each sentence; Third, extract and fuse the word-level language features of each word in each sentence. Fourth, the vectorized representation of each word is fused with word-level language features to obtain the text feature representation of each word; Fifth, determine whether the current word is the last word of the sentence. If it is, input the result into the sentence-level feature processing module; otherwise, return to step two. (2) The processing flow of the sentence-level feature processing module is as follows: First, obtain the text feature representation of each word in each sentence output by the word-level feature processing module; Second, fuse the text feature representation of each word in each sentence, and then calculate the semantic feature representation of the sentence according to formulas (1)-(3); Third, each sentence is vectorized to obtain its original vector representation, preserving the original features of the sentences; fourth, sentence-level linguistic features of each sentence are extracted and fused. Fifth, the sentence-level linguistic features, the original vector representation of the sentence, and the semantic feature representation of the sentence are fused together to obtain accurately fused sentence-level features; Sixth, based on the precisely fused sentence-level features and the three different weight matrices for sentence investigation, labeling, and start, the investigation vector, labeling vector, and result vector of each sentence are calculated using formulas (4)-(6); Seventh, based on the research vector, tag vector and result vector of each sentence, the attention optimization representation of each sentence is calculated by formula (7), and the attention optimization representation is used as the final text feature representation of each sentence; Eighth, determine if the current sentence is the last sentence of the comment. If so, input the result into the comment-level feature processing module; otherwise, return to the first step. (3) The processing flow of the comment-level feature processing module is as follows: First, obtain the text feature representations of all sentences in the comment output by the sentence-level feature processing module; Second, the text feature representations of all sentences in the comment are integrated, and then the text semantic feature representation of the comment is calculated according to formulas (8)-(10); Third, the comment text is vectorized to obtain the original vector representation of the comment, preserving the original features of the comment; fourth, the comment-level linguistic features of the comment are extracted and fused. Fifth, the comment-level linguistic features of the comment text, the original vector representation of the comment, and the semantic feature representation of the comment are fused to obtain a precisely fused comment-level feature; Sixth, based on the precisely fused comment-level features and the three different weight matrices of comment investigation, labeling, and start, the investigation vector, labeling vector, and result vector of the comment are calculated respectively by formulas (11)-(13); Seventh, based on the calculated comment investigation vector, labeling vector, and result vector, the multi-head attention optimization representation of the comment is calculated by formula (14), and this multi-head attention optimization representation is used as the final text feature representation of the comment; Eighth, input the results into the global feature processing module. (4) The processing flow of the global feature processing module is as follows: First, obtain the text feature representation of the comment output by the comment-level feature processing module; Second, extract and merge the user action features of the comment; Third, extract and merge the item features of the comment; Fourth, merge the text feature representation of the comment, the user action features of the comment, and the item features of the comment to obtain the accurately merged global feature of the comment. Fifth, based on the precisely fused global features of the comments, the final global depth feature representation of the comments is calculated using formulas (15)-(17); Sixth, the global deep feature representation of the comments is mapped to the feature-classification space to obtain the category prediction results of the comments.

2. The method for analyzing fake reviews based on multi-level and multi-dimensional feature fusion according to claim 1, characterized in that: The processing steps of the word-level feature processing module are as follows: Start at page 201; P202 Enter comment text; Page 203 performs sentence segmentation and word segmentation on the comment text; P204 embeds a vectorized representation of each word; P205 Extracts and merges word-level language features for each word; P206 integrates the vectorized representation of each word with word-level language features to obtain the text feature representation of each word; P207 checks if the current word is the last word of the sentence. If it is, proceed to P208 sequentially; otherwise, return to P204. P208 Input the results into the sentence-level feature processing module. Page 209, End.

3. The method for analyzing fake reviews based on multi-level and multi-dimensional feature fusion according to claim 1, characterized in that: The processing steps of the sentence-level feature processing module are as follows: Start at page 301; P302 Reads the text feature representation of each word in each sentence output by the word-level feature processing module; P303 integrates the text feature representation of each word in each sentence and calculates the semantic feature representation of each sentence according to formulas (1)-(3); P304 Vectorize each sentence to obtain the original vector representation of the sentence; P305 Extracts and fuses sentence-level language features for each sentence; P306 integrates the sentence-level linguistic features, the original vector representation of the sentence, and the semantic feature representation of the sentence to obtain accurately integrated sentence-level features; P307 Based on the precisely fused sentence-level features and formulas (4)-(6), calculate the single-sentence query vector, tag vector, and result vector of the comment; P308 Calculate the attention optimization representation of each sentence based on the single sentence query vector, tag vector and result vector, and formula (7), as the final text feature representation of each sentence; P309 Determine if the current sentence is the last sentence of the comment. If it is, proceed to P310 in sequence; otherwise, return to P302. P310 Input the results into the comment-level feature processing module. End of page 311.

4. The method for analyzing fake reviews by multi-level and multi-dimensional feature fusion according to claim 1, characterized in that: The processing steps of the comment-level feature processing module are as follows: Start at page 401; P402 Reads the text feature representations of all sentences output by the sentence-level feature processing module; P403 The text feature representations of all sentences in the comment are integrated, and the text semantic feature representation of the entire comment is calculated according to formulas (8)-(10); P404 Vectorizes the comment text to obtain the original vector representation of the comment; P405 Extract and merge comment-level linguistic features of comments; P406 integrates the comment-level linguistic features of the comment text, the original vector representation of the comment, and the semantic feature representation of the comment to obtain a precisely integrated comment-level feature; P407 Based on the precisely fused comment-level features and formulas (11)-(13), calculate the full-text search vector, tag vector, and result vector of the comment respectively; P408 Calculate the multi-head attention optimization representation of the comment based on the full-text search vector, tag vector and result vector of the comment, as well as formula (14), and use it as the final text feature representation of the comment; P409 Input the results into the global feature processing module. End of page 410.

5. The method for analyzing fake reviews by multi-level and multi-dimensional feature fusion according to claim 1, characterized in that: The processing steps of the global feature processing module are as follows: Start at page 501; P502 Reads the text feature representation of the comment output by the comment-level feature processing module; P503 Extracts and merges user action features from comments; P504 Extracts and merges item features from comments; P505 integrates the textual features of the comment, the user action features of the comment, and the item features of the comment to obtain a precisely integrated global feature of the comment; P506 The final global depth feature representation of the comments is calculated based on the precisely fused global features of the comments and formulas (15)-(17); P507 performs a feature-classification space mapping on the global deep feature representation of the comment to obtain the category prediction result of the comment. Page 508 is the end.

6. The method for analyzing fake reviews by multi-level and multi-dimensional feature fusion according to claim 1, characterized in that: The calculation formulas required for each module in the method described are defined as follows: (1) Formula for calculating the contextual semantic features of a single comment sentence （1） （2） （3） Where || represents the concatenation operation. (2) Formulas for calculating the single-sentence query vector, the single-sentence tag vector, and the single-sentence result vector. （4） (3) Calculation formula for optimizing the attention of a single comment sentence （7） (4) Formula for calculating the semantic features of the full text context of the comment （8） （9） （10） (5) Formulas for calculating the full-text search vector, the full-text tag vector, and the full-text result vector of the comments. （11） （12） （13） (6) Calculation formula for optimizing the full text attention of comments （14） (7) Global deep feature calculation formula （15） （16） （17）。

Images