E-commerce false comment recognition system based on neural network

Abstract

The invention discloses an e-commerce false comment recognition system based on a neural network in the technical field of false comment recognition, and the system comprises an information acquisition module, an information preprocessing module, a feature extraction module, a model construction module and a classification output module, and the model construction module is constructed based on a hierarchical neural network and a convolutional neural network. The model construction module is used for learning discourse representation of a comment text through a hierarchical neural network by utilizing word embedding and then combining multivariate discrete features through a convolutional neural network to generate a feature vector, and the classification output module is used for splicing a text semantic feature vector and the multivariate discrete feature vector. According to the method, the text content of the comments and the behavior features of the reviewers are subjected to feature vector extraction through the information acquisition module, the information preprocessing module, the feature extraction module, the model construction module and the classification output module, and then the extracted feature vectors are spliced and classified, so that the detection performance of the false comments is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of false review identification, and in particular is an e-commerce false review identification system based on neural network. Background Art

[0002] In recent years, with the rapid development and popularization of Internet technology and mobile terminals, people's consumption patterns have gradually changed. The rise of online shopping platforms and review websites has led more and more consumers to switch from offline shopping to online transactions. At the same time, a large number of Internet users share their consumption experience and evaluate the quality of products or services by posting comments on relevant product pages, and refer to other users' comments when making consumption decisions. At the same time, merchants can also use review information to timely understand customers' consumption needs, improve their own product quality and service level, and promote the healthy development of merchants.

[0003] In order to create a good reputation and earn more profits, some unscrupulous businesses will use monetary rewards as an inducement to write a large number of false reviews by hiring online writers, "brushing orders", and offering cash back for positive reviews. On the one hand, they write false positive reviews for their own products or services to drive user discussions, build a good reputation, stimulate user consumption, and thus increase sales and profits; on the other hand, they write a large number of false negative reviews for competitors to discredit and suppress them. These false review activities will reduce consumers' shopping experience and are not conducive to the stable development of e-commerce platforms. Therefore, technical means are needed to identify false reviews.

[0004] Previous studies have shown that due to people's bias towards truth and the misuse of false review identification markers, the accuracy of manual false review identification is only slightly higher than 50%. When manually judging the authenticity of review content, there is a lack of information such as body language, facial expressions, tone of voice, etc. as in face-to-face communication. In addition, in order to prevent consumers from easily identifying the falsity of the reviews, false reviews are always written very realistically. In general, manual false review identification is very difficult and inaccurate.

[0005] At present, researchers mainly conduct research on false review identification by using the text content of the review and the behavioral characteristics of the reviewer respectively. Due to the differences in data types between the review text and the behavioral characteristics of the reviewer, in-depth research is needed on how to use these two features simultaneously. In addition, most existing studies select classification features based on traditional experience, which is subjective and has low interpretability. Using a large number of classification features for model prediction may contain noise information, thus affecting model performance. Therefore, it is necessary to propose an e-commerce false review identification system based on neural networks. Summary of the invention

[0006] In order to solve the above problems, the purpose of the present invention is to provide an e-commerce false review identification system based on neural network, which extracts feature vectors of the text content of the comments and the behavioral characteristics of the commentators through information acquisition module, information preprocessing module, feature extraction module, model building module and classification output module, and then splices and classifies the extracted feature vectors, thereby improving the detection performance of false reviews.

[0007] In order to achieve the above-mentioned purpose, the technical solution of the present invention is as follows: an e-commerce false review identification system based on neural network, including an information acquisition module, an information preprocessing module, a feature extraction module, a model building module and a classification output module.

[0008] The information acquisition module is used to collect product reviews when users browse product reviews on web pages and convert the product reviews into corpus data.

[0009] The information preprocessing module is used to preprocess the corpus data converted by the information acquisition module according to the preprocessing process, and process the corpus data into a standardized text data format.

[0010] The feature extraction module is used to establish a corpus database based on existing corpus data, train according to the characteristics of the corpus data marked as true or false comments, and then output potential clues for false comment identification as text language information, other language feature information and non-language feature information, extract word vectors from the text language information and improve the word vectors; and extract features from other language feature information and non-language feature information.

[0011] The model building module is built based on hierarchical neural networks and convolutional neural networks. The model building module is used to use word embedding to learn the text semantic feature vector of the review text through the hierarchical neural network, and to generate a multivariate discrete feature vector by combining multivariate discrete features through the convolutional neural network.

[0012] The classification output module is used to concatenate and classify the text semantic feature vector and the multivariate discrete feature vector and then output the false comment detection results.

[0013] The principle of the basic solution is: the information acquisition module is responsible for grabbing comment data from the product comment page of the e-commerce platform and converting it into corpus data for subsequent processing. This step is the basis of data collection and ensures the accuracy and comprehensiveness of subsequent analysis.

[0014] Preprocessing is the process of data cleaning and standardization. The information preprocessing module aims to convert raw text into a standardized text data format to facilitate subsequent feature extraction and model training.

[0015] The feature extraction module is based on previous research and expert experience, combined with the characteristics of the fake review annotation dataset, to extract key features from text language information, other language feature information, and non-language feature information. For text language information, word embedding technology is used to convert words into vector representations, and then improve them to better capture the semantic relationship between words.

[0016] The model building module is based on hierarchical neural networks (such as recurrent neural networks (RNNs) or their variants LSTM, GRU) and convolutional neural networks (CNNs). Hierarchical neural networks are used to learn deep semantic representations of review texts, while convolutional neural networks are used to process multivariate discrete features (such as non-linguistic features) and generate feature vectors.

[0017] The feature vectors are concatenated and classified through the classification output module to output the detection results of fake comments. This step realizes the conversion from features to final decisions.

[0018] The beneficial effects of the basic program are:

[0019] 1. By combining text semantic features and non-linguistic features, the system can more comprehensively capture the characteristics of fake reviews and effectively improve the recognition accuracy of fake reviews. Compared with traditional manual recognition or recognition methods based only on text content, this system has significant advantages in identifying complex and hidden fake reviews.

[0020] 2. By utilizing the powerful learning ability of neural networks, the system can automatically learn the underlying patterns of false reviews from large amounts of data, reducing reliance on manually labeled data and enhancing the generalization ability of the model.

[0021] 3. The automated processing flow greatly shortens the detection time of false reviews and improves processing efficiency, enabling e-commerce platforms to respond to false review issues more quickly and maintain a good market environment and user experience.

[0022] Furthermore, the preprocessing process includes filtering illegal characters, word segmentation, English word processing, and removing stop words.

[0023] The beneficial effect of the basic scheme is that by filtering illegal characters and removing stop words, the proportion of noisy data can be significantly reduced, the quality and purity of the data can be improved, and more accurate and reliable input can be provided for subsequent feature extraction and model training.

[0024] Word segmentation and English word processing help to cut the text into independent vocabulary units and reduce the diversity of vocabulary, so that the feature vector can more accurately represent the semantic information of the text. Removing stop words can reduce the dimension of the feature vector and reduce the complexity of the model, while retaining the vocabulary that contributes significantly to the semantics of the text and improving the generalization ability of the model.

[0025] Preprocessed text data can better reflect the semantic features of the text, thereby improving the accuracy of false review identification. Preprocessing can also reduce overfitting during model training and improve the stability and robustness of the model.

[0026] Furthermore, the feature extraction module includes a word vector extraction unit, a language feature extraction unit and a non-language feature extraction unit.

[0027] The word vector extraction unit is used to obtain the word vector in the text language information, model the relationship between the context and the target word through a neural network, use the corpus database for training, map each word to a 200-dimensional vector space, and determine the relationship between two words by taking the cosine method. Then, the Labeled-LDA model is used to integrate the information in the data to calculate the importance of the word under each topic tag.

[0028] The language feature extraction unit is used to extract part-of-speech features, personal pronoun ratios, readability indicators, and sentiment tendency language features.

[0029] The non-linguistic feature extraction unit is used to extract non-linguistic features from the review rating, rating deviation, user activity, merchant review filtering rate, review support and sentiment consistency.

[0030] The beneficial effects of the basic scheme are: word vectors can capture the semantic relationship between words, making similar words closer in the vector space, thereby improving the ability to represent text semantics. The Labeled-LDA model can identify topics in text and calculate the importance of each word under the topic, which helps to identify possible topic manipulation behaviors in fake reviews.

[0031] Part-of-speech features and the proportion of personal pronouns can reflect the language style of the text and help distinguish real reviews from fake ones. Sentimental features can reflect the emotional color of the text. False reviews may deliberately exaggerate or minimize emotional tendencies to influence readers' judgment.

[0032] Features such as user activity and comment support can reflect user behavior patterns and help identify whether there is malicious or false review behavior.

[0033] Characteristics such as merchant review filtering rate can reflect the merchant's credibility. False reviews are often filtered or deleted by merchants, resulting in abnormal review filtering rate.

[0034] Furthermore, the word vector extraction unit uses the Yelp public dataset to pre-train English word vectors, and adopts the Chinese word vector data open sourced by Tencent AI Lab.

[0035] The beneficial effects of the basic solution are: The Yelp public dataset is a large-scale English dataset that contains rich user review information. Using this dataset for pre-training English word vectors can capture the semantic relationship and contextual information between English words, thereby generating high-quality English word vectors.

[0036] The Chinese word vector data open sourced by Tencent AI Lab contains more than 8 million Chinese words, each of which corresponds to a 200-dimensional vector. These data have significantly improved coverage, freshness, and accuracy, and can cover a variety of vocabulary types, including phrases, new Internet words, etc. These data can be used to generate high-quality Chinese word vectors.

[0037] By using both English and Chinese word embedding data, the model can learn the commonalities and differences between different languages, thereby enhancing the ability of cross-language processing. This is of great significance for natural language processing tasks in a multilingual environment.

[0038] Furthermore, word embedding refers to mapping each word into a high-dimensional, continuous, real-valued vector space using a word vector extraction unit to vectorize the word representation.

[0039] The beneficial effect of the basic scheme is that word embedding technology can convert discrete words into continuous vector representations, so that semantically similar words are also close in distance in the vector space. This representation method helps computers better understand and process semantic relationships in language, thereby improving the accuracy of natural language processing tasks.

[0040] Word embedding technology enables the model to better understand the meaning of words by learning contextual information, thereby alleviating the problems caused by contextual ambiguity. This ability enables the model to make reasonable inferences and representations based on the context when faced with new and unseen words.

[0041] Furthermore, the multivariate discrete features refer to other language features and non-language features extracted by the language feature extraction unit and the non-language feature extraction unit.

[0042] The beneficial effects of the basic scheme are: multi-discrete features cover language features such as part-of-speech features, personal pronoun ratio, readability index, sentiment tendency, and non-linguistic features such as review ratings, rating deviations, user activity, merchant review filtering rate, review support, and sentiment consistency. These features can describe the text from multiple dimensions, enhancing the comprehensiveness and accuracy of feature representation. By taking these features as input, the model can understand the text content more comprehensively, thereby more accurately identifying false reviews.

[0043] Furthermore, the model building module also includes an attention mechanism unit, which is used to introduce an attention mechanism in the model building module, use a scoring function score to score the importance of each word in the sentence representation, and enhance the attention mechanism unit's ability to capture key information by calculating the importance weight of each word in the sentence.

[0044] The beneficial effect of the basic scheme is that the attention mechanism allows the model to dynamically assign different importance weights to each input element when processing the input sequence, so that the model can pay more attention to the information related to the current task. This selective focus on key information helps improve the overall performance of the model.

[0045] The model can generate attention weight distributions when processing input sequences, which reflect the model's judgment of the importance of each input element. This helps us understand how the model makes decisions, thereby enhancing the interpretability of the model.

[0046] Furthermore, the attention mechanism unit obtains the attention probability distribution vector by calculating the output of the convolutional neural network and the initial state of all words, and then outputs the weighted text semantic features.

[0047] The beneficial effect of the basic solution is that the attention mechanism unit can dynamically calculate the importance weight of each word in the text based on the output of the convolutional neural network and the initial state of the word, thereby focusing on key information. This dynamic focusing method helps the model understand the semantic content of the text more accurately and improves the accuracy of text semantic understanding.

[0048] Furthermore, the classification output module includes a classification unit and an output unit, and the classification unit is used to concatenate and classify the text semantic feature vector and the multivariate discrete feature vector in the model building module.

[0049] The output unit is used to output the concatenated text semantic feature vector and multivariate discrete feature vector as a false comment or a true comment.

[0050] The beneficial effect of the basic scheme is that the classification unit concatenates the text semantic feature vector with the multivariate discrete feature vector to achieve comprehensive feature fusion. This fusion method not only retains the deep semantic information in the text, but also introduces multi-dimensional discrete features such as user behavior and comment time, thereby greatly enhancing the classification ability of the model.

[0051] By integrating multiple features, the model can more comprehensively capture the key information in the comments and reduce classification errors caused by missing information or noise interference. This directly improves the accuracy of classification and enables the system to more accurately identify fake comments.

[0052] Furthermore, the classification output module also includes a result display unit, which is used to display the false comment detection results to the user in a visual manner.

[0053] The beneficial effects of the basic solution are: the result display unit displays the detection results in a visual way, allowing users to understand at a glance which comments are judged to be false comments. This intuitiveness greatly improves the user's reading efficiency and comprehension ability. Through the visual display, users can intuitively see how the system makes judgments based on the content of the comments and other features. This visual verification method helps to improve users' credibility of the system results and reduce misunderstandings and doubts. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of the structure of a neural network-based e-commerce false review identification system in an embodiment of the present invention.

[0055] Figure 2 Schematic diagram of the structure of the feature extraction module of the neural network-based e-commerce false review identification system in an embodiment of the present invention.

[0056] Figure 3 Schematic diagram of the structure of the classification output module of the neural network-based e-commerce false review identification system in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following is further described in detail through specific implementation methods:

[0058] Example:

[0059] Basically as attached Figure 1 -Attached Figure 3 As shown: A neural network-based e-commerce fake review identification system includes an information acquisition module, an information preprocessing module, a feature extraction module, a model building module and a classification output module.

[0060] The information acquisition module is used to collect product reviews when users browse product reviews on web pages and convert the product reviews into corpus data.

[0061] The information preprocessing module is used to preprocess the corpus data converted by the information acquisition module according to the preprocessing process and process the corpus data into a standardized text data format. The preprocessing process includes filtering illegal characters, word segmentation, English word processing and removing stop words.

[0062] The feature extraction module is used to establish a corpus database based on existing corpus data, train according to the characteristics of the corpus data marked as true or false comments, and then output potential clues for false comment identification as text language information, other language feature information and non-language feature information, extract word vectors from the text language information and improve the word vectors; and extract features from other language feature information and non-language feature information.

[0063] The feature extraction module includes a word vector extraction unit, a language feature extraction unit and a non-language feature extraction unit.

[0064] The word vector extraction unit is used to obtain word vectors in text language information, model the relationship between context and target words through neural networks, and use corpus databases for training. Each word is mapped to a 200-dimensional vector space, and the relationship between two words is determined by cosine. Then, the Labeled-LDA model is used to integrate the information in the data to calculate the importance of the word under each topic label. The word vector extraction unit uses the Yelp public data set to pre-train English word vectors, and uses the open source Chinese word vector data from Tencent AI Lab.

[0065] The language feature extraction unit is used to extract part-of-speech features, personal pronoun ratios, readability indicators, and sentiment tendency language features.

[0066] The non-linguistic feature extraction unit is used to extract non-linguistic features from the review rating, rating deviation, user activity, merchant review filtering rate, review support and sentiment consistency.

[0067] The model building module is built based on hierarchical neural networks and convolutional neural networks. The model building module is used to learn the text semantic feature vector of the review text through the hierarchical neural network using word embedding, and to generate a multivariate discrete feature vector by combining multivariate discrete features through the convolutional neural network. Word embedding refers to mapping each word into a high-dimensional, continuous real-valued vector space using a word vector extraction unit to vectorize the word representation. Multivariate discrete features refer to other language features and non-language features extracted by the language feature extraction unit and the non-language feature extraction unit.

[0068] The model building module also includes an attention mechanism unit, which is used to introduce the attention mechanism into the model building module, use the scoring function score to score the importance of each word in the sentence representation, and enhance the attention mechanism unit's ability to capture key information by calculating the importance weight of each word in the sentence. The attention mechanism unit calculates the output of the convolutional neural network and the initial state of all words to obtain the attention probability distribution vector, and then outputs the weighted text semantic features.

[0069] The classification output module is used to splice and classify the text semantic feature vector and the multi - variable discrete feature vector, and then output the false review detection result.

[0070] The classification output module includes a classification unit and an output unit. The classification unit is used to splice and classify the text semantic feature vector and the multi - variable discrete feature vector in the model construction module.

[0071] The output unit is used to output the spliced text semantic feature vector and multi - variable discrete feature vector as false reviews or true reviews.

[0072] The classification output module also includes a result display unit, which is used to display the false review detection result to the user in a visual way.

[0073] The specific implementation process is as follows: First, the information acquisition module collects data when the user browses the product, including product review data and merchant's own data, etc. Then the data pre - processing module pre - processes the collected data. The data pre - processing module removes illegal characters or meaningless characters in the review text, such as punctuation marks, special characters, garbled characters, etc. Then the continuous character sequence is segmented into words according to rules. For Chinese text, word segmentation needs to be carried out according to Chinese semantics. For English text, the words in the English text are uniformly made lowercase, and the original forms of words in different tenses are restored through stemming. Remove words with high frequency of occurrence in the text but meaningless for analysis and research, such as "a, this, is, 了, 是", etc.

[0074] After the data pre - processing module finishes processing the data, the vector features of the data will be extracted through the feature extraction module. The relationship between the context and the target word is modeled through the word vector extraction unit. The Yelp public dataset is used for the pre - training of English word vectors, and the data open - sourced by Tencent AI Lab is used for Chinese word vectors.

[0075] Then the Labeled - LDA model is used to integrate the class information in the data, and the importance degree of words under each topic label is calculated. The obtained word vectors integrating Labeled - LDA are beneficial to making the expression of word vectors more abundant.

[0076] In addition to the text language information, other language features are extracted through the language feature extraction unit, such as part - of - speech features, the proportion of other personal pronouns, readability indicators, and sentiment tendencies, etc. The non - language feature extraction unit extracts features from aspects such as the rating of the review, rating deviation, user activity, merchant review filtering rate, review support degree, sentiment consistency, etc., to enrich the evaluation indicators of the model.

[0077] After extracting word vector features, other language features, and non-language features, the fake review monitoring model can be constructed through the model building module. Through the attention mechanism unit, the importance of each word in the sentence is scored using a scoring function to obtain an attention probability distribution vector. The distribution vector is multiplied and accumulated with the hidden layer output of the historical node to obtain the final text semantic features, thereby obtaining a better semantic expression effect.

[0078] The quantitative indicators of other language features and non-language features are discrete values, so the feature dictionary is used to represent the discrete feature value corresponding to each comment. Then, a convolutional neural network is used to convert the multivariate discrete feature matrix into a low-dimensional continuous feature vector.

[0079] The fake review monitoring model uses word embedding to learn text semantic feature representation through a hierarchical neural network, and then uses a convolutional neural network to combine multiple discrete features to generate a feature vector.

[0080] Finally, the text semantic features and multivariate discrete feature vectors are input into the classification output module, and the classification unit in the classification output module is used to splice the text semantic features and the multivariate discrete feature vectors. Then, the output unit is used to output the spliced ​​text semantic feature vectors and the multivariate discrete feature vectors as false comments or real comments. Finally, the result display unit is used to display the false comment detection results to the user in a visual manner.

[0081] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0082] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A neural network-based e-commerce fake review identification system, characterized by: It includes information acquisition module, information preprocessing module, feature extraction module, model building module and classification output module; The information acquisition module is used to collect product reviews when users browse product reviews on the web page and convert the product reviews into corpus data; An information preprocessing module is used to preprocess the corpus data converted by the information acquisition module according to a preprocessing process, and process the corpus data into a standardized text data format; The feature extraction module is used to establish a corpus database based on the existing corpus data, train according to the characteristics of the corpus data marked as true or false comments, and then output the potential clues for false comment identification as text language information, other language feature information and non-language feature information, extract word vectors from the text language information and improve the word vectors; extract features from other language feature information and non-language feature information; The model building module is built based on a hierarchical neural network and a convolutional neural network. The model building module is used to learn the text semantic feature vector of the review text through a hierarchical neural network using word embedding, and to generate a multivariate discrete feature vector by combining multivariate discrete features through a convolutional neural network; The classification output module is used to concatenate and classify the text semantic feature vector and the multivariate discrete feature vector and then output the false comment detection results.

2. The e-commerce false review identification system based on neural network according to claim 1, characterized in that: The preprocessing process includes filtering illegal characters, word segmentation, English word processing, and removing stop words.

3. The e-commerce false review identification system based on neural network according to claim 2 is characterized by: The feature extraction module includes a word vector extraction unit, a language feature extraction unit and a non-language feature extraction unit; The word vector extraction unit is used to obtain the word vector in the text language information, model the relationship between the context and the target word through a neural network, use the corpus database for training, map each word to a 200-dimensional vector space, and determine the relationship between two words by taking the cosine method. Then, the Labeled-LDA model is used to integrate the information in the data to calculate the importance of the word under each topic label; A language feature extraction unit, used to extract part-of-speech features, personal pronoun ratio, readability index and sentiment tendency language features; The non-linguistic feature extraction unit is used to extract non-linguistic features from the review rating, rating deviation, user activity, merchant review filtering rate, review support and sentiment consistency.

4. The e-commerce false review identification system based on neural network according to claim 3 is characterized by: The word vector extraction unit uses the Yelp public dataset to pre-train English word vectors, and adopts the Chinese word vector data open sourced by Tencent AILab.

5. The e-commerce false review identification system based on neural network according to claim 4 is characterized by: Word embedding refers to the use of word vector extraction units to map each word into a high-dimensional, continuous real-valued vector space to vectorize the word representation.

6. The e-commerce false review identification system based on neural network according to claim 5, characterized in that: Multivariate discrete features refer to other language features and non-language features extracted by the language feature extraction unit and the non-language feature extraction unit.

7. The e-commerce false review identification system based on neural network according to claim 6, characterized in that: The model building module also includes an attention mechanism unit; The attention mechanism unit is used to introduce the attention mechanism into the model building module, and uses the scoring function score to score the importance of each word in the sentence representation. By calculating the importance weight of each word in the sentence, the attention mechanism unit's ability to capture key information is enhanced.

8. The e-commerce false review identification system based on neural network according to claim 7, characterized in that: The attention mechanism unit obtains the attention probability distribution vector by calculating the output of the convolutional neural network and the initial state of all words, and then outputs the weighted text semantic features.

9. The e-commerce false review identification system based on neural network according to claim 8, characterized in that: The classification output module includes a classification unit and an output unit; The classification unit is used to concatenate and classify the text semantic feature vector and the multivariate discrete feature vector in the model building module; The output unit is used to output the concatenated text semantic feature vector and multivariate discrete feature vector as a false comment or a true comment.

10. The e-commerce false review identification system based on neural network according to claim 9, characterized in that: The classification output module also includes a result display unit; The result display unit is used to display the false review detection results to the user in a visual manner.

Images