Mobile e-commerce false review detection method and system based on large language model

Abstract

The present application relates to the technical field of natural language processing and data analysis, in particular to a mobile e-commerce false comment detection method and system based on a large language model, which generates false comment text by designing prompt words, constructs a training data set, extracts five types of numerical features, then classifies and describes the text according to the numerical feature value range based on the set feature rules, converts the numerical features into text natural language features, obtains a rule set, further constructs a rule data set, fine-tunes the large language model using the rule data set, obtains a first large language model, and distills the first large language model to a second large language model with smaller model parameters, which can capture the language style, word usage habits and expression methods specific to the mobile e-commerce field, better distinguish real feedback and false comments, and improve the accuracy of false comment identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing and data analysis, in particular to a mobile e-commerce fake review detection method and system based on a large language model. BACKGROUND

[0002] Consumer reviews of goods or services on different e-commerce platforms have become an important factor affecting purchase intention. Merchants' pursuit of good reviews, and the use of fake reviews by some unscrupulous merchants or competitors to hype or attack products have made the problem of fake reviews on online platforms more serious.

[0003] A large number of research and technical means have been invested in the detection of fake reviews. Technically, it mainly relies on natural language processing and machine learning techniques, which can be identified by combining artificial feature design with classification algorithms. With the rapid development of deep learning technology, the large language model is applied to the mobile e-commerce platform to capture the deep semantic features of the text and establish the relationship between the text and the associated entities, and to identify the fake content that has been deeply forged, which provides the possibility for fine and semantic mobile e-commerce review analysis.

[0004] However, the existing methods often have the following limitations: first, the existing detection methods mainly screen reviews through keyword matching or pre-defined rule libraries, such as marking reviews containing obvious promotional words or frequently repeating derogatory language. However, as the forms of fake reviews become more flexible and complex, this method is difficult to cover a wide range of review expressions, and it is even more difficult to deal with ambiguous language or deliberately changed text patterns in reviews; second, there is a lack of understanding of complex semantics, and it is difficult to accurately judge the authenticity of reviews when facing reviews with deep meanings or emotional tendencies; third, there is a lack of customized learning ability for specific fields. At present, many detection methods mainly rely on models trained on large-scale general data sets, and these models often fail to capture the unique language style, word habit and expression of the field when facing specific fields such as e-commerce, catering and tourism. General models may miss these key field features when generalizing, thereby affecting the accuracy of distinguishing fake reviews from real feedback. In actual detection, it may show a high misjudgment rate, affecting the authenticity of review identification.

[0005] Therefore, there is an urgent need to invent a large language model detection method that is specifically designed for mobile e-commerce environment, can accurately identify fake reviews, and has high efficiency and scalability. SUMMARY

[0006] To solve the above problems in the background art, the present application provides a mobile e-commerce fake review detection method and system based on a large language model.

[0007] The technical scheme of the present application is as follows:

[0008] A mobile e-commerce false comment detection method based on a large language model, comprising the following steps:

[0009] S1, obtaining real comment text of a user on a product, generating false comment text based on a designed prompt word, integrating data of the real comment text and the false comment text, and constructing a training data set;

[0010] S2, sequentially performing data cleaning and data deduplication on the data in the training data set to obtain standard data;

[0011] S3, sequentially performing sentence segmentation and word segmentation on the standard data based on semantic paragraphs, obtaining a plurality of sentences and words, extracting numerical features based on the plurality of sentences and words, the numerical features including sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-tag ratio features;

[0012] S4, based on the set feature rules, the numerical features are classified and described according to the feature value range they are in, a plurality of feature texts are obtained, and the plurality of feature texts are combined to obtain a rule set;

[0013] The training data set and the rule set are data fused to construct a rule data set;

[0014] S5, using the rule data set to fine-tune train the large language model to obtain a first large language model;

[0015] S6, distilling the first large language model to a second large language model with smaller model parameters, and performing false comment detection on new product comment text.

[0016] Specifically, the extraction of the text readability feature in S3 is based on the total number of words, the total number of sentences, the total number of syllables, and the total number of words, considering complex words and professional terms to obtain a readability score, which is represented by the following formula:

[0017]

[0018] Wherein, C complex is the penalty coefficient of complex words, C technical is the penalty coefficient of professional terms,

[0019] The design of the prompt word in S1 includes role prompt words, task prompt words, and prompt word feature constraints.

[0020] Further, the role prompt word is: you are a net shopping review generation assistant; the task prompt word is: please generate a number of reviews based on product information, the product information includes: product category, product name, price, product characteristics;

[0021] The prompt word feature constraint is: the generated review needs to meet the following characteristics: the review length is within a certain range; the review contains exaggerated or unrealistic use experience; simulate the writing habits of real users and contain colloquial expressions; the review contains rating tendency and emotional intensity.

[0022] The prompt word based on the design generates false review text, and each false review text needs to contain the following elements: purchase experience, use feeling, specific detail description, while paying attention to the length of the review; each false review needs to consider: word diversity, emotional expression intensity, detail description specific degree, and authenticity degree.

[0023] The S6 in the distillation process:

[0024] The first probability distribution of the first large language model output about the real review text or the false review text is used as a soft label;

[0025] The soft label is used as the target label for training the second large language model, and the second large language model is trained using the training data set to output a second probability distribution;

[0026] The KL divergence loss is calculated based on the first probability distribution and the second probability distribution, the cross-entropy loss is calculated based on the second probability distribution and the target label, and the model parameter of the second large language model is updated based on the KL divergence loss and the cross-entropy loss.

[0027] The training data set D constructed in the S1 text is represented as follows:

[0028] D text = real review text S1{user published text content, true or false label, publishing time}+false review text S2{generated text content, true or false label, publishing time}.

[0029] The S5 uses a rule data set to fine-tune the large language model, injects a low-rank adapter into the large language model, updates the rank parameter of the low-rank adapter according to the gradient value of the loss function, and uses the loss function: Where y i is the output corresponding to the i-th text sample, P is the probability function, M fine-tuned is the first large language model after fine-tuning training, and N is the batch size.

[0030] The data cleaning in S2 is: filtering the emoticons and garbled characters in the data, further unifying the full-width or half-width characters in the data, sequentially performing the conversion between simplified Chinese and traditional Chinese, normalizing the case, and removing the comments with less than 10 words.

[0031] The application also provides a mobile e-commerce false comment detection system based on a large language model, comprising:

[0032] The data acquisition and generation module is used for acquiring real comment texts of users on commodities, generating false comment texts based on designed prompt words, integrating real comment texts and false comment texts, and constructing a training data set;

[0033] The data processing module is used for sequentially performing data cleaning and data deduplication on the data in the training data set, and obtaining standard data;

[0034] The feature extraction module is used for sequentially performing sentence segmentation and word segmentation on the standard data based on semantic paragraphs, obtaining a plurality of sentences and words, extracting numerical features based on the plurality of sentences and words, and the numerical features include: sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-tag ratio features;

[0035] The feature conversion module is used for performing text classification description on the numerical features according to their corresponding feature value ranges based on the set feature rules, obtaining a plurality of feature texts, combining the plurality of feature texts, obtaining a rule set, and performing data fusion on the training data set and the rule set to construct a rule data set;

[0036] The model fine-tuning module is used for fine-tuning training the large language model using the rule data set to obtain a first large language model;

[0037] The false comment detection module is used for distilling the first large language model to a second large language model with smaller model parameters, and performing false comment detection on new commodity comments.

[0038] The application has the following beneficial effects:

[0039] 1、The application generates false comment texts by designing prompt words, constructs a training data set, extracts five types of numerical features including sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-tag ratio features, then performs text classification description on the numerical features according to their corresponding feature value ranges based on the set feature rules, converts the numerical features into text natural language features, obtains a rule set, further constructs a rule data set, can capture the language style, word habit and expression method specific to the mobile e-commerce field, better distinguishes real feedback and false comments, and improves the accuracy of false comment identification.

[0040] 2. The use of rule data sets for fine-tuning training of large language models obtains a first large language model, and distills the first large language model to a second large language model with smaller model parameters, which can maintain good results without using numerical features, and ensures efficient detection in the mobile e-commerce environment. DETAILED DESCRIPTION

[0041] Exemplary embodiments of the present disclosure will be described in greater detail below.

[0042] Embodiments

[0043] The present embodiment provides a mobile e-commerce fake review detection method based on a large language model, which includes the following steps:

[0044] S1, obtaining real review text of users on commodities, generating fake review text based on designed prompt words, and integrating data of real review text and fake review text to construct a training data set.

[0045] In the present embodiment, the real review text comes from public review data of actual e-commerce platforms. Under the premise of complying with relevant laws and platform rules, the review text of users on various commodities is obtained. The user reviews in the product detail page are obtained by using the BeautifulSoup library in Python.

[0046] Further, fake review text is generated based on designed prompt words. The design of prompt words should meet the generation of a large number of diversified online shopping review data, and the content should be as close as possible to the characteristics of artificial writing.

[0047] The design of prompt words includes role prompt words, task prompt words, and prompt word feature constraints. The role prompt words are: you are an online shopping review generation assistant; the task prompt words are: please generate a number of reviews based on product information, the product information includes: product category, product name, price, and product features.

[0048] The prompt word feature constraints are: the generated reviews need to meet the following features: (1) the review length is within a certain range, for example, the review length is set to be between 20-100 words; (2) the review contains exaggerated or unrealistic use experience; (3) simulate the writing habits of real users, including colloquial expressions; (4) the review contains rating tendency and emotional intensity.

[0049] At the same time, each fake review text needs to contain the following elements: purchase experience, use feeling, and specific detail description, while paying attention to the length of the review; each fake review needs to consider: word diversity, emotional expression intensity, detail description specificity, and authenticity.

[0050] Taking a digital product as an example, the prompt words can be designed as:

[0051] Role prompt: You are a professional digital product review generation assistant.

[0052] Task prompt: Based on the following product information, generate 10 fake reviews. Product information includes: product category: wireless earphones; product name: True Wireless Noise-Canceling Earphones Pro; price: 1299 yuan; product features: active noise cancellation, 40-hour battery life, Bluetooth 5.3, Hi-Res certification.

[0053] Prompt features constraints: (1) Review length: 30-80 characters; (2) Reviews contain fictional and unreasonable use scenarios, exaggerated product performance, and unrealistic comparisons; (3) Simulate real user writing habits, including colloquial expressions; (4) Reviews contain rating tendencies and emotional intensity, with 80% positive reviews and 20% negative reviews.

[0054] Reviews must include: specific use scenarios, technical parameters, subjective experience descriptions, and comparisons with other brands. Time range: January 1, 2023 to present.

[0055] Further, the real review text and the generated fake review text are fused to construct the training data set D text , represented as follows:

[0056] D text = Real review text S1 {user's text content, true or false label, publication time} + fake review text S2 {generated text content, true or false label, publication time}.

[0057] S2, the data in the training data set is sequentially cleaned and de-duplicated to obtain standard data.

[0058] The main purpose of step S2 is to clean and de-duplicate the data in the training data set to filter out valid data. The data cleaning process is as follows: filter out emoticons, garbled characters, unify full-width or half-width characters, convert between simplified and traditional Chinese, standardize case, and remove reviews with fewer than 10 words.

[0059] The data de-duplication process is as follows: convert the cleaned data into text vectors, calculate the similarity between text vectors based on cosine similarity, set the similarity value greater than or equal to the set similarity threshold as a duplicate review, and only keep the review with the earlier publication time. In addition, comments that do not contain any substantive evaluation content are marked as "invalid comments" and removed. Finally, the standard data is obtained.

[0060] S3, sequentially performing sentence division and word division on the standard data based on semantic paragraphs, to obtain a plurality of sentences and words, and extracting numerical features based on the plurality of sentences and words, the numerical features including: sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-tag ratio features.

[0061] In step S3, the standard data is divided into sentences based on semantic paragraphs, and long texts are split into a plurality of sentences, so as to facilitate subsequent individual statistics of the part-of-speech of each sentence, and facilitate subsequent statistics of the absence rate of verbs or other parts of speech of each sentence.

[0062] Then, word division is performed on each sentence to obtain a continuous word sequence. On the basis of word division, each word obtained after word division is assigned a corresponding part-of-speech label. For example, for the sentence "this mobile phone performance is good", after word division, "this", "mobile phone", "performance", and "good" are obtained, and each word is then assigned a corresponding part-of-speech label, such as DT (determiner), Q (quantifier), N (noun), V (verb), and V (adjective). Finally, the following is obtained: this / DT, mobile phone / N, performance / N, and good / V.

[0063] The main purpose of step S3 is to extract numerical features based on a plurality of sentences and words. The numerical features of the present application include: sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-tag ratio features. Each numerical feature will be introduced one by one as follows:

[0064] For sentiment polarity features, the existing large-scale Chinese pre-training RoBERTa model is used to extract sentiment polarity features from the words obtained after word division, and a sigmoid function is used to map the sentiment polarity features to the interval [0, 1] to obtain a sentiment polarity score.

[0065] For text readability features, traditional text readability is mainly evaluated based on sentence length and average word length of each sentence. The present application adds processing of complex words and professional terms to more accurately evaluate the readability of Chinese text and obtain a readability score, which is expressed by the following formula:

[0066]

[0067] wherein C complex is a penalty coefficient of complex words, C technical is a penalty coefficient of professional terms,

[0068] Among them, complex words are defined as words containing ≥3 characters and uncommon combinations, which have a negative impact on the readability of the text, so a penalty coefficient for complex words is added when calculating. The penalty coefficient of complex words represents the proportion of complex words in the total number of words. The higher the proportion, the worse the readability of the text, and the larger the penalty coefficient.

[0069] Considering the impact of professional terms, a mobile e-commerce professional term library is defined to identify and mark domain-related terms. If the comment contains words outside the mobile e-commerce professional term library, it will be considered as an external word, and its impact on the readability of the text will be calculated through the penalty coefficient of professional terms. Texts with a high proportion of external words usually contain too many professional terms, making it difficult for ordinary readers to understand. According to the calculated readability score, if the score is less than 30, the text is judged to be low readability. The setting of 30 takes into account the user's reading comprehension ability. If the text is difficult to understand, its readability is poor.

[0070] For part-of-speech distribution features, in this embodiment, we mainly calculate the noun density, adjective concentration, verb-object structure missing rate, text topic distribution features, and type-tag ratio indicators. The specific expressions are as follows: Part-of-speech distribution features

[0071] (1) Noun density is the proportion of nouns in a sentence, which is calculated by dividing the number of nouns in a sentence by the total number of words. The calculation formula is:

[0072] Noun density reflects the frequency of noun use in a sentence. In general, texts with higher noun density may be more descriptive.

[0073] (2) Adjective concentration represents the concentration of adjectives in a sentence, measuring the frequency of adjectives appearing in a sentence. The calculation formula is:

[0074] Adjective concentration reflects the strength of evaluation or description of the text. Sentences with high adjective concentration often contain more subjective color and emotional expression.

[0075] (3) Verb-object structure missing rate: Verb-object structure refers to a phrase composed of a verb and a noun. The completeness of the verb-object structure is closely related to the meaning of the sentence. Texts with high verb-object structure missing rate may have incomplete grammar or unclear expression.

[0076] The calculation formula is:

[0077] For text topic distribution features, a topic probability distribution vector is generated for each comment. The LDA model generates a topic probability distribution vector for each comment, which is represented by the following formula:

[0078]

[0079] where p i is the probability that review d belongs to the i-th topic, and K is the number of topics.

[0080] Then the topic dispersion entropy H is calculated:

[0081] In order to make the result in the interval [0, 1], the normalized dispersion entropy H is defined as: norm

[0082]

[0083] where H norm = 0 represents complete concentration on a single topic, and H norm = 1 represents a completely uniform topic distribution.

[0084] When the value is a low dispersion entropy (H norm <0.5), it indicates that the review is of a small number of topics, and is inclined to be judged as a fake review feature.

[0085] The type-token ratio (TTR) is an indicator used to measure the lexical diversity of a text. It reflects the use of different words in a text and can help analyze the richness of the language. Fake reviews have limited vocabulary richness and often have a lower TTR value.

[0086] The basic formula of TTR is:

[0087] where V is the number of different words in the text, and N is the total number of words in the text.

[0088] S4, based on the set of feature rules, the numerical features are classified and described according to the corresponding feature value range, obtaining a number of feature texts, and combining a number of feature texts to obtain a rule set.

[0089] The training data set and the rule set are fused to construct a rule data set.

[0090] In step S4, the numerical features are converted into more interpretable text classification descriptions by setting the feature rules, obtaining a number of feature texts.

[0091] For numerical features: sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-token ratio features, the feature rules are set as follows:

[0092] ​(1) Sentiment polarity feature: Sentiment polarity score range [0.9-1.0]: "extremely strong emotional expression"; [0.7-0.89]: "strong emotional tendency"; [0.4-0.69]: "medium emotional intensity"; [0.0-0.39]: "weak or contradictory emotional expression".

[0093] (2) Text readability feature: Readability score <30: "low readability"; Readability score range [30-60]: "medium readability"; Readability score >60: "high readability".

[0094] (3) Text topic distribution feature: Discreteness entropy value <0.3: "highly concentrated on a single topic"; Discreteness entropy value range [0.3-0.5]: "moderate topic concentration"; Discreteness entropy value >0.5: "topic dispersion.

[0095] (4) Type-token ratio feature (TTR): TTR <0.4: "high lexical repetition rate"; TTR range 0.4-0.6: "medium lexical diversity"; TTR >0.6: "rich vocabulary.

[0096] (5) Part-of-speech distribution feature: Adjective proportion >40%: "exaggerated description"; Adjective proportion 20%-40%: "regular description"; Adjective proportion <20%: "lack of subjective evaluation.

[0097] For example, for the review text "This phone is amazing, the photos are super clear, and the battery life is particularly good!", after the numerical feature extraction in step S3, the obtained numerical features are:

[0098] Sentiment polarity feature: 0.95, Text readability feature: 93.7, Part-of-speech distribution feature: Adjective proportion 40%, Text topic distribution feature: Discreteness entropy value 0.3, Type-token ratio feature: 0.6.

[0099] In this example, too much numerical feature data is not conducive to subsequent large language model understanding of the internal meaning of the numerical value, so the numerical feature is described in natural language according to its feature value range, obtaining:

[0100] Sentiment polarity feature: The review shows extremely strong positive emotion; Expression characteristics: The text is clear and easy to understand, the content is highly concentrated on product advantages, and the language is natural and fluent; Language mode: Adjective usage frequency is significantly higher than ordinary reviews, and praise vocabulary density is larger.

[0101] Then combine several feature texts to obtain a rule set. Perform data fusion on the training data set and the rule set to construct a rule data set, so that the large language model can receive text and feature information simultaneously when processing.

[0102] S5, fine-tuning training the large language model using the rule dataset to obtain a first large language model.

[0103] The role of step S5 is to adapt the large language model to the data of a specific field by fine-tuning it. The large language model used in this embodiment is Qwen2.5-72B. By injecting a low-rank adapter into the large language model, fine-tuning training the large language model using the rule dataset, updating the rank parameter of the low-rank adapter according to the gradient value of the loss function until the preset number of iterations is reached, and the fine-tuning training is completed. Define the forward propagation as y = M Qwen2.5 (x) + A T · B T · x, A, B ∈ R d×r , B ∈ R r×d , where M Qwen2.5 represents the large language model used, x represents the vector input to the large language model, A and B represent matrices of a specific dimension, T represents the transpose of the matrix, d represents the dimension of the hidden layer of the large language model, r represents the rank of the low-rank adapter, and r < d to ensure the low-rank property and reduce computational complexity. In this invention, the value of r is specified in the range of 400-1200.

[0104] During fine-tuning training, define the fine-tuning task as discriminating fake reviews in the training dataset, update the rank parameter of the low-rank adapter according to the gradient value of the loss function, and use the loss function as: where y i is the output corresponding to the i-th text sample, P is the probability function, M fine-tuned is the first large language model after fine-tuning training, and N is the batch size. In each iteration period, the rank parameter of the low-rank adapter is updated until the preset number of iterations is reached, the fine-tuning training is completed, and the first large language model is obtained.

[0105] At the same time, the prepared verification data is used to evaluate the large language model, and the accuracy, precision, recall and F1 score are calculated. The accuracy is used to measure the proportion of all reviews that are correctly classified, the precision is the proportion of actual fake reviews in the set of fake reviews predicted, the recall is the proportion of actual fake reviews that are correctly predicted, and the F1 score is used to comprehensively measure the detection ability and accuracy of the large language model for fake reviews, to evaluate the current training progress.

[0106] S6, distill the first large language model to a second large language model with smaller model parameters, and perform fake review detection on new product review text.

[0107] The first large language model is distilled to a second large language model with smaller model parameters, the Qwen2.5-1.5B with smaller parameters is used as the second large language model, which is suitable for various small business scenarios and has smaller resource utilization. In the distillation process, the second large language model can only accept text comments as input and make the most accurate judgment without using features by learning the knowledge of the first large language model.

[0108] In the distillation process: the first probability distribution output by the first large language model about the real review text or the fake review text is used as a soft label; the soft label is used as a target label for training the second large language model, and the second large language model is trained using the training data set to output a second probability distribution; the KL divergence loss is calculated based on the first probability distribution and the second probability distribution, the cross-entropy loss is calculated based on the second probability distribution and the target label, and the model parameter of the second large language model is updated based on the KL divergence loss and the cross-entropy loss, which is expressed as follows:

[0109] Loss=α·CrossEntropyLoss(P student ,Y true )+(1-α)·KLDivLoss(P student ,P teacher ),

[0110] Wherein, Loss is the loss function of the model parameter update of the second large language model, α is a weight coefficient, P student is the second probability distribution, Y true is the target label, P teacher is the first probability distribution, CrossEntropyLoss is the cross-entropy loss, and KLDivLoss is the KL divergence loss.

[0111] Finally, the second large language model is actually deployed for mobile e-commerce environment, and the new commodity comment text is detected for fake review to obtain the detection result of the comment being true or false.

[0112] In this embodiment, the mobile e-commerce fake review detection method based on the large language model used by the application is compared with other SVM, BERT and Few-shot three different methods, and the comparison results are shown in Table 1.

[0113] Table 1 Mobile e-commerce fake review detection results

[0114] Method Accuracy Precision Recall F1 Score SVM 71.3％ 74.6％ 70.2％ 72.3％ BERT 82.5％ 80.3％ 76.6％ 78.4％ Few-shot 58.7％ 57.1％ 53.8％ 55.4％ The invention 87.1％ 85.8％ 83.4％ 84.6％

[0115] As can be seen from Table 1, the accuracy of the traditional machine learning method Few-shot is the lowest, only 58.7%, and the performance of SVM is also poor, with an accuracy of 71.3%, which shows that the detection method without fine-tuning training still has great limitations in identifying fake reviews. The performance of BERT is better than SVM and Few-shot, reaching an accuracy of 82.5%. And the present application can better perform the task of detecting fake e-commerce reviews by performing data processing, feature extraction, feature conversion and model fine-tuning, with an accuracy of 87.1% in detecting fake reviews, and the precision, recall and F1 score evaluation indicators are all better than other methods, so the present application can better perform the task of detecting fake e-commerce reviews.

[0116] The present application also provides a mobile e-commerce fake review detection system based on a large language model, comprising:

[0117] A data acquisition and generation module is used to acquire real review text of users on commodities, generate fake review text based on designed prompt words, integrate data of real review text and fake review text, and construct a training data set;

[0118] A data processing module is used to sequentially perform data cleaning and data deduplication on data in the training data set to obtain standard data;

[0119] A feature extraction module is used to sequentially perform sentence segmentation and word segmentation on the standard data based on semantic paragraphs, to obtain a plurality of sentences and words, and to extract numerical features based on the plurality of sentences and words, wherein the numerical features include sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-tag ratio features;

[0120] A feature conversion module is used to perform text classification description on the numerical features according to their feature value ranges based on set feature rules, to obtain a plurality of feature texts, to combine the plurality of feature texts to obtain a rule set, and to perform data fusion on the training data set and the rule set to construct a rule data set;

[0121] A model fine-tuning module is used to fine-tune the large language model using the rule data set to obtain a first large language model;

[0122] A fake review detection module is used to distill the first large language model to a second large language model with smaller model parameters, and to perform fake review detection on new commodity reviews.

Claims

1. A method for detecting fake reviews in mobile commerce based on a large language model, characterized in that, The method comprises the following steps: S1, obtaining real review text of a user on a commodity, generating false review text based on a designed prompt word, integrating data of the real review text and the false review text, and constructing a training data set; S2, sequentially performing data cleaning and data deduplication on the data in the training data set to obtain standard data; S3, sequentially performing sentence segmentation and word segmentation on the standard data based on semantic paragraphs, obtaining a plurality of sentences and words, extracting numerical features based on the plurality of sentences and words, the numerical features including emotional polarity features, text readability features, part-of-speech distribution features, text theme distribution features, and type-tag ratio features; S4, classifying and describing the numerical features according to their feature value ranges based on a set of feature rules, obtaining a plurality of feature texts, and combining the plurality of feature texts to obtain a rule set; The training data set and the rule set are fused to construct a rule data set; S5, using the rule data set to fine-tune a large language model to obtain a first large language model; S6, distilling the first large language model to a second large language model with smaller model parameters, and detecting false reviews of new commodity review texts.

2. The mobile e-commerce fake review detection method based on a large language model according to claim 1, characterized in that, The extraction of the text readability features in S3 considers complex vocabulary and professional terms based on total word count, total sentence count, total phoneme count, and total word count to obtain a readability score, which is represented by the following formula: wherein C complex is a penalty coefficient for complex vocabulary, C technical penalty coefficient for professional terms, 3.The mobile e-commerce fake review detection method based on a large language model according to claim 1, characterized in that, The design of the prompt word in S1 includes role prompt words, task prompt words, and prompt word feature constraints.

4. The mobile e-commerce fake review detection method based on a large language model according to claim 3, characterized in that, The role prompt word is: You are a net shopping review generation assistant. The task prompt word is: Please generate a plurality of reviews based on product information, the product information including product category, product name, price, and product features. The prompt word feature constraint is: The generated reviews need to meet the following features: The review length is within a certain range; the review contains exaggerated or unrealistic use experience; it simulates the writing habits of real users and contains colloquial expressions; the review contains rating tendency and emotional intensity.

5. The mobile e-commerce fake review detection method based on a large language model according to claim 3, characterized in that, The false review text generated based on the designed prompt word contains the following elements: purchase experience, use feeling, and specific detail description, while paying attention to the length of the review; each false review needs to consider the following: word diversity, emotional expression intensity, detail description specificity, and authenticity. 6.The mobile e-commerce fake review detection method based on a large language model according to claim 1, characterized in that, In the distillation process in S6: The first probability distribution of the first large language model output on the real review text or the false review text is used as a soft label; The soft label is used as a target label for training the second large language model, and the second large language model is trained using the training data set to output a second probability distribution; The KL divergence loss is calculated based on the first probability distribution and the second probability distribution, the cross-entropy loss is calculated based on the second probability distribution and the target label, and the model parameter of the second large language model is updated based on the KL divergence loss and the cross-entropy loss.

7. The mobile e-commerce fake review detection method based on a large language model according to claim 1, characterized in that, The training data set D constructed in S1 text is represented as follows: D text = real review text S1 {text content posted by a user, true or false label, posting time} + fake review text S2 {generated text content, true or false label, posting time}. 8.The mobile e-commerce fake review detection method based on a large language model according to claim 1, characterized in that, In S5, the rule data set is used to fine-tune the large language model to inject a low-rank adapter into the large language model, and the rank parameter of the low-rank adapter is updated according to the gradient value of the loss function, and the loss function used is: where y i is the output corresponding to the i-th text sample, P is the probability function, M fine-tuned is the fine-tuned first large language model, and N is the batch size. 9.The mobile e-commerce fake review detection method based on large language model according to claim 1, characterized in that, The data cleaning in S2 is: filtering the emoticons and garbled characters in the data, further unifying the full-width or half-width characters in the data, sequentially converting the simplified and traditional Chinese, normalizing the case, and removing the comments with less than 10 words.

10. A mobile e-commerce fake review detection system based on a large language model, characterized in that, Comprise: A data acquisition and generation module is configured to acquire real review texts of users on commodities, generate false review texts based on designed prompt words, integrate the real review texts and the false review texts, and construct a training data set; A data processing module is configured to sequentially perform data cleaning and data deduplication on the data in the training data set to obtain standard data; A feature extraction module is configured to sequentially perform sentence segmentation and word segmentation on the standard data based on semantic paragraphs to obtain a plurality of sentences and words, and extract numerical features based on the plurality of sentences and words, wherein the numerical features include: sentiment polarity features, text readability features, part-of-speech distribution features, text topic distribution features, and type-tag ratio features; A feature conversion module is configured to perform text classification description on the numerical features according to their corresponding feature value ranges based on a set of feature rules to obtain a plurality of feature texts, combine the plurality of feature texts to obtain a rule set, and perform data fusion on the training data set and the rule set to construct a rule data set; A model fine-tuning module is configured to fine-tune a large language model using the rule data set to obtain a first large language model; A false review detection module is configured to distill the first large language model to a second large language model with smaller model parameters, and perform false review detection on new commodity reviews.

Images