A personalized review recommendation method and system based on implicit dimension mining
By constructing the M3E-base-TextDimension model and rewriting comments using a large language model, and combining user needs to filter comment dimensions, the problem of personalized recommendations among massive amounts of comments was solved, achieving high-quality personalized comment recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2024-01-24
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies struggle to provide personalized recommendations from the vast amount of online product reviews of varying quality, especially in accurately matching the implicit dimensions of reviews to the needs of new users.
By constructing the M3E-base-TextDimension model, we use a large language model to identify entities in comments and rewrite the comments. We then combine user needs to select the appropriate comment dimensions, generate comment and demand embeddings, and calculate cosine similarity for Top-N recommendations.
It enables high-quality rewriting of comment text and implicit dimension mining, expands the dimension set of comments, improves the accuracy and efficiency of personalized recommendations, and meets the personalized needs of new users.
Smart Images
Figure CN117932151B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of natural language processing and artificial intelligence, and in particular to a personalized comment recommendation method and system based on implicit dimension mining. Background Technology
[0002] Online product reviews are a crucial component of e-commerce websites, reflecting the genuine experiences of past users. They not only help new users quickly understand the quality of products and services but also help merchants identify product strengths and weaknesses. Any positive or negative opinions expressed in these online product reviews are known as online word-of-mouth, and research has shown the impact of online reviews on new users' purchase intentions. However, the development of the internet has led to an explosive growth in review data, with varying review quality. Therefore, recommending useful information from this massive volume of reviews is of great significance to users, merchants, and platforms alike.
[0003] In recent years, review recommendation technologies have focused on the task of predicting review usefulness. A key step in this task is review feature extraction. Among these features, text content is the most frequently used, primarily including parts-of-speech tags, textual content, and readability. However, text-based usefulness prediction techniques only reflect the general usefulness of reviews and are difficult to personalize recommendations based on the individual needs of new users. Therefore, researchers have developed review recommendation technologies based on features other than text content, including non-textual content, product-related factors, and reviewer-related factors. Summary of the Invention
[0004] Given that existing methods for automatically extracting implicit features from comments and new user needs struggle to simultaneously guarantee accuracy and reduce human intervention, and that implicit features themselves lack good inductive properties, making it difficult to intuitively match new user needs, this invention is proposed.
[0005] Therefore, the problem this invention aims to solve is how to utilize large language models to mine implicit dimensions in comments and new user needs in order to achieve more accurate personalized comment recommendations.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0007] In a first aspect, embodiments of the present invention provide a personalized review recommendation method based on implicit dimension mining, which includes: constructing an M3E-base-TextDimension model by filtering online product reviews and fine-tuning the M3E-base model; identifying entities in the reviews using a large language model and rewriting the reviews to extract review dimensions; combining user needs and review dimensions, filtering a set of review dimensions that meet user needs as demand dimensions using the large language model; inputting the review dimensions and demand dimensions into the M3E-base-TextDimension model to generate review embeddings and demand embeddings; and performing Top-N review recommendations by calculating the cosine similarity between the review embeddings and demand embeddings.
[0008] As a preferred embodiment of the personalized review recommendation method based on implicit dimension mining described in this invention, the construction of the M3E-base-TextDimension model includes the following steps: using Python's pandas library to deduplicate online product review data; filtering out garbled text by calculating the word segmentation rate (WSR) and performing word segmentation using the jieba library; in the web crawler program, length filtering of the text data to determine the practical length of the text data; using the HanLP toolkit for Chinese word segmentation and filtering the text data by constructing a stop word list; selecting entity words using the PKU part-of-speech tagging set to determine candidate dimensions for text analysis; and fine-tuning the M3E-base model using a dataset containing review text and candidate dimensions to construct the M3E-base-TextDimension model.
[0009] As a preferred embodiment of the personalized comment recommendation method based on implicit dimension mining described in this invention, the method for filtering text data by constructing a stop word list includes the following steps: importing the HanLP toolkit and checking if an existing stop word list exists; if it exists, directly loading the existing stop word list; if it does not exist, constructing a custom stop word list; trying various word segmentation algorithms and evaluating their effects to select the optimal algorithm for word segmentation of the text data; determining whether it is necessary to filter stop words; if so, traversing the word segmentation results and matching them with the stop word list to remove stop words, while outputting the filtered word segmentation results to obtain a list of keywords without stop words; if not, directly outputting the word segmentation results.
[0010] As a preferred embodiment of the personalized review recommendation method based on implicit dimension mining described in this invention, the fine-tuning of the M3E-base model using a dataset containing review text and candidate dimensions includes the following steps: constructing a fine-tuning dataset for M3E-base based on review text and candidate dimensions, including training data, validation data, and test data; determining whether the dataset file format is two columns including the original review, the original review, and the candidate dimension; if it is two columns, loading the M3E-base model, using the two columns of data as text input, and setting their similarity label to 1; if the first column of data is combined with the second column of data from other samples, setting their similarity label to 0; and fine-tuning the M3E-base model using the cross-entropy loss function based on the labeled samples to obtain the M3E-base-TextDimension model.
[0011] As a preferred embodiment of the personalized review recommendation method based on implicit dimension mining described in this invention, the method comprises: constructing an entity recognition Prompt to identify product attributes, functions, and defect entities in the original review text; constructing an entity recognition text rewriting Prompt to guide the ChatGLM-Pro model in rewriting the original review text based on the entity recognition results, and converting implicit dimensions into explicit dimensions while ensuring semantic integrity; performing word segmentation and part-of-speech tagging on the newly generated rewritten text using HanLP, retaining only words with parts of speech n, nz, and nx as review dimensions; constructing a dimension checking Prompt to determine whether the review dimension is mentioned in the original review using a large language model; if the dimension is not mentioned in the original review, it is deleted as an irrelevant dimension; if the dimension is mentioned in the original review, it is added to the review dimension set.
[0012] As a preferred embodiment of the personalized comment recommendation method based on implicit dimension mining described in this invention, the following steps are included in selecting comment dimensions that meet user needs as demand dimensions using a large language model: Constructing a user demand Prompt, including user demand statements and a candidate set of comment dimensions, to dynamically generate comment dimensions that meet different user needs; using the user demand statements and the candidate set of comment dimensions as input to guide the large language model to determine whether each candidate dimension is relevant to the user demand statement; the determination method is to construct a Prompt template for determining relevance, fill the Prompt template with user needs and the comment dimensions to be judged, and perform relevance judgment through the large language model; for comment dimensions that the large language model determines are relevant to user needs, they are retained as demand dimensions, and semantic enhancement of the user demand text is performed using the demand dimensions.
[0013] As a preferred embodiment of the personalized review recommendation method based on implicit dimension mining described in this invention, the generation of review embeddings and demand embeddings includes the following steps: loading the M3E-base-TextDimension model; for each review text, directly concatenating the review dimension corresponding to the review text to the end of the review text to obtain text in the form of review and review dimension; for each user demand text, directly concatenating the demand dimension obtained in step S3 to the end of the demand text to obtain text in the form of demand and demand dimension; merging all review and review dimension texts, as well as demand and demand dimension texts, into a list, and inputting the text list line by line into the M3E-base-TextDimension model; taking the output text embeddings to form a matrix, wherein the order is consistent with the input text order, and generating separate review embeddings and demand embeddings according to the embedding matrix; and saving the review embeddings and demand embeddings.
[0014] Secondly, embodiments of the present invention provide a personalized review recommendation system based on implicit dimension mining, which includes a fine-tuning module for constructing an M3E-base-TextDimension model by filtering online product reviews and fine-tuning the M3E-base model; a review dimension module for identifying entities in reviews using a large language model and rewriting reviews to extract review dimensions; a demand dimension module for combining user needs and review dimensions and filtering a set of review dimensions that meet user needs as demand dimensions using a large language model; a generation module for inputting review dimensions and demand dimensions into the M3E-base-TextDimension model to generate review embeddings and demand embeddings; and a recommendation module for performing Top-N review recommendations by calculating the cosine similarity between review embeddings and demand embeddings.
[0015] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of the personalized review recommendation method based on implicit dimension mining as described in the first aspect of the present invention.
[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the personalized review recommendation method based on implicit dimension mining as described in the first aspect of the present invention.
[0017] The beneficial effects of this invention are as follows: By constructing a custom stop word list and utilizing multiple open-source stop word libraries, this invention achieves effective filtering of text data, improving the quality of subsequent text processing; leveraging the powerful semantic understanding and generation capabilities of large language models, it achieves high-quality comment text rewriting and implicit dimension mining, expanding the dimension set of comments; utilizing the Prompt method effectively utilizes the capabilities of large models while ensuring the controllability of model generation behavior, achieving precise filtering of comment and demand dimensions; and for dimensional information, it constructs a new data type of "text + dimension," employing a method of first merging text and then encoding it into embeddings to further improve model performance. Addressing the issues of short user demand texts and implicit dimensions in comment texts, we use a method of supplementing dimensions to perform horizontal and vertical semantic enhancement. Given that existing embedding models are mostly trained on natural language text data, we construct a dataset of "text + dimension" type data and fine-tune the model to allow the model to better handle this type of data. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating the steps of a personalized review recommendation method based on implicit dimension mining.
[0020] Figure 2 This is a schematic diagram of the framework structure of a personalized review recommendation method based on implicit dimension mining. Detailed Implementation
[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0024] Example 1
[0025] Reference Figures 1-2 This is the first embodiment of the present invention, which provides a personalized review recommendation method based on implicit dimension mining, including:
[0026] S1: By filtering online product reviews and fine-tuning the M3E-base model, a system is built.
[0027] M3E-base-TextDimension model.
[0028] Specifically, it includes the following steps:
[0029] S1.1: Use Python's pandas library to deduplicate online product review data.
[0030] S1.2: Filter out garbled text by calculating the word segmentation rate (wsr) and use the jieba library for word segmentation.
[0031] Specifically, the difference in word segmentation rate is used to filter garbled text. This difference refers to the fact that normal text has a segmentation rate greater than 2, while garbled text exhibits a segmentation rate close to 1. By observing this difference in segmentation rate, a preliminary judgment of garbled text can be made based on the numerical value of the segmentation rate. The specific formula for the WSR (Wide Segmentation Rate) is as follows:
[0032] wsr = w / s
[0033] Where wsr represents the word segmentation rate, w represents the sentence length (i.e., the number of characters), and s represents the number of words.
[0034] Furthermore, the jieba library was used for word segmentation. Tests showed that data with a segmentation rate of less than 1.55% contained the vast majority of garbled text, with the remaining small amount of garbled text being manually filtered.
[0035] S1.3: In the web crawler program, the text data is filtered by length to determine the practical length of the text data.
[0036] S1.4: Use the HanLP toolkit for Chinese word segmentation and filter the text data by building a stop word list.
[0037] Specifically, import the HanLP toolkit and check if an existing stop word list exists. If it exists, load the existing stop word list directly; otherwise, construct a custom stop word list. Try various word segmentation algorithms and evaluate their effects to select the optimal algorithm for word segmentation of the text data. Determine whether stop words need to be filtered. If so, iterate through the word segmentation results and match them with the stop word list to remove stop words. Output the filtered word segmentation results to obtain a list of keywords without stop words. If not, output the word segmentation results directly.
[0038] It should be noted that the stop word list includes both Chinese and English words, and its main purpose is to filter out words that have little impact on the semantics of the text. Considering the high degree of freedom in comment texts, which often contain a mixture of Chinese and English, this invention compiles several well-known open-source stop word libraries, such as those from Baidu, Harbin Institute of Technology, and NLTK, as the final stop word resource.
[0039] S1.5: Select entity words using the PKU part-of-speech tagging set to determine candidate dimensions for text analysis.
[0040] Specifically, the PKU part-of-speech tagging set is used as the standard. Based on user needs, the entity attention is reflected by three parts of speech: nouns (n), foreign characters (nz), and other proper nouns (nx). Therefore, this invention only retains words of the above three parts of speech as candidate dimensions.
[0041] It should be noted that the benefit of adding entity recognition to text rewriting is to achieve more granular implicit dimension mining.
[0042] S1.6: Fine-tune the M3E-base model using a dataset containing comment text and candidate dimensions to build the M3E-base-TextDimension model.
[0043] It should be noted that the fine-tuning dataset for M3E-base, constructed based on comment text and candidate dimensions, includes training data, validation data, and test data. The dataset file format is determined to be two columns: the original comment, the original comment, and the candidate dimension. If it is two columns, the M3E-base model is loaded, and the two columns of data are used as text input, with their similarity label set to 1 (representing high semantic relevance). If the first column of data is combined with the second column of data from other samples, their similarity label is set to 0 (representing semantic irrelevance). Based on the above positive and negative samples, the M3E-base model is fine-tuned using the cross-entropy loss function to obtain the M3E-base-TextDimension model.
[0044] S2: Use a large language model to identify entities in the comments and rewrite the comments to extract comment dimensions.
[0045] Preferably, considering that the ability of a large language model to complete entity recognition tasks directly affects the quality of subsequent text rewriting, this invention selects a model with hundreds of billions of parameters, such as ChatGPM-Pro, as the base, and utilizes its powerful text generation and semantic understanding capabilities to achieve high-quality text rewriting in order to uncover implicit dimensions.
[0046] Furthermore, an entity recognition prompt is constructed to identify entities such as product attributes, functions, and defects in the original review text. Based on the identified entities, a text rewriting prompt is built to guide the ChatGPM-Pro model in rewriting the original review text, converting implicit dimensions into explicit dimensions while maintaining basic semantic integrity. For the newly generated rewritten text, the M3E-base-TextDimension model is used to extract candidate dimensions, which are then checked and removed. The remaining candidate dimensions are used as review dimensions. A dimension checking prompt is constructed, using a large language model to determine whether the review dimension is mentioned in the original review. If the dimension is not mentioned in the original review, it is deleted as an irrelevant dimension.
[0047] S3: Combining user needs and comment dimensions, a set of key dimensions is selected through a large language model to confirm the demand dimensions.
[0048] Furthermore, the user requirement Prompt is constructed by including user requirement statements (recording the user's specific needs) and a candidate set of comment dimensions to dynamically generate comment dimensions that meet different user needs. The user requirement statements and the candidate set of comment dimensions are used as input to guide the large language model to determine whether each candidate dimension is relevant to the user requirement statement. The determination method is to construct a Prompt template for determining relevance, fill the Prompt template with the user requirement and the comment dimension to be judged, and then use the large language model to determine the relevance. For comment dimensions that the large language model determines are relevant to the user requirement, they are retained as requirement dimensions, and the semantics of the user requirement text are enhanced using the requirement dimensions.
[0049] S4: Input the comment dimension and demand dimension into the M3E-base-TextDimension model to generate comment embedding and demand embedding.
[0050] Specifically, load the M3E-base-TextDimension model; for each comment text, directly append the comment dimension corresponding to the comment text to the end of the comment text to obtain text in the form of "comment + comment dimension"; for each user request text, directly append the request dimension obtained in step S3 to the end of the request text to obtain text in the form of "request + request dimension"; merge all "comment + comment dimension" texts and "request + request dimension" texts into a list, and input the text list line by line.
[0051] The M3E-base-TextDimension model is used to extract the output text embeddings into a matrix, where the order is consistent with the input text order. Separate comment embeddings and demand embeddings are generated based on the embedding matrix. The comment embeddings and demand embeddings are then saved.
[0052] Preferably, by merging and then encoding, vertical semantic enhancement of comment text and horizontal semantic enhancement of user needs are achieved, ultimately resulting in enhanced comment embedding and demand embedding.
[0053] S5: Top-N comment recommendations are made by calculating the cosine similarity between comment embeddings and demand embeddings.
[0054] Specifically, the semantically enhanced comment embedding and demand embedding saved in step S4 are loaded; the cosine similarity between each demand embedding and all other comment embeddings is calculated; the cosine similarity results are associated with the corresponding comments, and the comments are sorted in descending order of similarity; for each demand, the top N comments are taken as the recommendation results; and the N recommended comments corresponding to each demand are output, forming the correspondence between demand and Top-N recommended comments.
[0055] The preferred formula for calculating cosine similarity is as follows:
[0056]
[0057] Where u represents the demand embedding, r i Let sim(u,r) represent the i-th semantically enhanced comment. i ) represents u and r i The similarity is represented by θ.
[0058] It should be noted that, since text similarity focuses more on semantic similarity, this invention chooses cosine similarity. The cosine similarity is calculated between the semantically enhanced requirement embedding obtained by merging "requirement + requirement dimension" into M3E-base-TextDimension encoding and the semantically enhanced comment embedding obtained by merging "original comment + comment dimension" into M3E-base-TextDimension encoding.
[0059] Furthermore, this embodiment also provides a personalized review recommendation system based on implicit dimension mining, including a fine-tuning module for constructing an M3E-base-TextDimension model by filtering online product reviews and fine-tuning the M3E-base model; a review dimension module for identifying entities in reviews using a large language model and rewriting reviews to extract review dimensions; a demand dimension module for combining user needs and review dimensions and filtering key dimension sets through a large language model to confirm demand dimensions; a generation module for inputting review dimensions and demand dimensions into the M3E-base-TextDimension model to generate review embeddings and demand embeddings; and a recommendation module for performing Top-N review recommendations by calculating the cosine similarity between review embeddings and demand embeddings.
[0060] This embodiment also provides a computer device applicable to the personalized review recommendation method based on implicit dimension mining, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the personalized review recommendation method based on implicit dimension mining as proposed in the above embodiment.
[0061] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0062] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the personalized comment recommendation method based on implicit dimension mining as proposed in the above embodiments.
[0063] In summary, this invention achieves effective filtering of text data and improves the quality of subsequent text processing by constructing a custom stop word list and utilizing multiple open-source stop word libraries; it leverages the powerful semantic understanding and generation capabilities of large language models to achieve high-quality comment text rewriting and implicit dimension mining, expanding the dimension set of comments; and it effectively utilizes the capabilities of large models while ensuring the controllability of model generation behavior through the Prompt method, achieving precise selection of comment dimensions and demand dimensions.
[0064] Example 2
[0065] Reference Figures 1-2 This is the second embodiment of the present invention, which provides a personalized review recommendation method based on implicit dimension mining. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.
[0066] Specifically, over 400,000 rich Chinese product reviews were collected from public websites. After data cleaning and deduplication, 90,000 training samples, 2,000 validation samples, and 880 test samples were selected. The HanLP toolkit was used to segment and label the reviews.
[0067] Furthermore, a large language model is used to identify, rewrite, and remove redundant dimensions from the comment text, resulting in a high-quality comment dimension set that includes both the original explicit dimensions of the comment and the implicit dimensions mined after rewriting. Based on the new requirement statement input by the user, a Prompt is constructed to allow the large language model to automatically determine the relevance between the requirement and each comment dimension, and to filter out personalized requirement dimensions that match the user's interests. This dimension set is determined by the user's requirements.
[0068] Furthermore, the "comment text + corresponding comment dimension" and the "requirement statement + requirement dimension" are respectively input into the pre-trained semantic analysis model M3E-base-TextDimension to learn semantically enhanced comment expressions and user requirement expressions. The cosine similarity is calculated for all pre-processed semantically enhanced comment expressions and newly input user requirement expressions, and the Top-N comments that are closest to the semantics of the requirement are returned as the recommendation results.
[0069] Furthermore, the model's recommendation advantages were compared with other methods on the Amazon review matching dataset, with the test results on the Amazon-MRHP dataset shown in Table 1.
[0070] Table 1 Test results of the Amazon-MRHP dataset
[0071]
[0072] Wherein, PRR-LI indicates the M3E-base version without fine-tuning, PRR-LI-FT indicates the M3E-base version with fine-tuning, BiMPN, EG-CNN, EG-CNN, and PRHNet indicate text-only models, SSE-Cross, D&R Net, and MCR indicate multimodal models, Text-only indicates text-only models, Multimodal indicates multimodal models, Ours indicates the model of this invention, Clothing indicates clothing category data, Electronics indicates electronic product category data, Home indicates home furnishing product category data, M@5 indicates the MAP value of the first 5 comments, N@3 indicates the NDCG value of the first 3 comments, and N@5 indicates the NDCG value of the first 5 comments.
[0073] Preferably, as shown in Table 1, the model of the present invention performs competitively on this dataset. Using only text data, it achieves results that surpass or approach those of a multimodal model in some metrics; and after fine-tuning, the model's performance has been further improved.
[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A personalized review recommendation method based on implicit dimensionality mining, characterized in that: include, By filtering online product reviews and fine-tuning the M3E-base model, an M3E-base-TextDimension model is constructed. Utilize a large language model to identify entities in comments and rewrite the comments to extract comment dimensions; By combining user needs and comment dimensions, a set of comment dimensions that meet user needs is selected as the demand dimensions through a large language model. Input the comment dimension and demand dimension into the M3E-base-TextDimension model to generate comment embedding and demand embedding; Comment recommendations are made by calculating the cosine similarity between comment embeddings and demand embeddings. The construction of the M3E-base-TextDimension model includes the following steps: Use Python's pandas library to remove duplicates from online product review data; By calculating the word segmentation rate Filter out garbled text and use the jieba library for word segmentation; In a web crawler, the length of text data is filtered to determine its practical length. We used the HanLP toolkit for Chinese word segmentation and filtered the text data by building a stop word list. pass Part-of-speech tagging sets are used to select entity words to determine candidate dimensions for text analysis; Fine-tune the M3E-base model using a dataset containing comment text and candidate dimensions to build the M3E-base-TextDimension model; The fine-tuning of the M3E-base model using a dataset containing comment text and candidate dimensions includes the following steps: The fine-tuning dataset for M3E-base, constructed based on comment text and candidate dimensions, includes training data, validation data, and test data. The fine-tuning dataset contains two columns of data: the original comments and the combination of the original comments and candidate dimensions. Load the M3E-base model, take the two columns of data as text input, and set the similarity label to 1 when the first column of data and the second column of data come from the same original comment; When the first column of data is combined with the second column of data from other samples, the similarity label is set to 0; Based on the labeled samples, the M3E-base model is fine-tuned using the cross-entropy loss function to obtain the M3E-base-TextDimension model.
2. The personalized review recommendation method based on implicit dimension mining as described in claim 1, characterized in that: The process of filtering text data by constructing a stop word list includes the following steps: Import the HanLP toolkit and check if an existing stop word list exists. If the existing stop word list exists, load it directly; otherwise, construct a custom stop word list. Try various word segmentation algorithms and evaluate their effects to select the optimal algorithm for word segmentation of text data; Determine whether stop words need to be filtered. If so, iterate through the word segmentation results and match them with the stop word list to remove stop words. At the same time, output the filtered word segmentation results to obtain a list of keywords without stop words. If stop word filtering is not required, the word segmentation results will be output directly.
3. The personalized review recommendation method based on implicit dimension mining as described in claim 1, characterized in that: The process of rewriting comments to extract comment dimensions includes the following steps: Construct an entity recognition prompt, and use the entity recognition prompt to identify product attributes, functions, and defect entities in the original review text; Construct an entity recognition text rewriting Prompt to guide the ChatGLM-Pro model to rewrite the original comment text based on entity recognition results, and transform implicit dimensions into explicit dimensions while ensuring semantic integrity. The newly generated text is segmented and tagged with parts of speech using HanLP, and only words with parts of speech n, nz, and nx are retained as comment dimensions. Build a dimension check prompt and use a large language model to determine whether the comment dimension is mentioned in the original comment; If a dimension is not mentioned in the original comment, it will be deleted as an irrelevant dimension. If a dimension is mentioned in the original comment, it will be added to the comment dimension set.
4. The personalized review recommendation method based on implicit dimension mining as described in claim 1, characterized in that: The process of selecting comment dimensions that meet user needs as demand dimensions using a large language model includes the following steps: The user requirement Prompt is constructed by including user requirement statements and a candidate set of comment dimensions, in order to dynamically generate comment dimensions that meet different user requirements; The user demand statement and the candidate set of comment dimensions are used as input to guide the large language model to determine whether each candidate dimension is related to the user demand statement. The method involves constructing a Prompt template to determine relevance, filling the template with user needs and the dimensions of the comments to be judged, and then using a large language model to determine relevance. For comment dimensions that the large language model determines are related to user needs, they are retained as demand dimensions, and semantic enhancement of user demand text is performed using demand dimensions.
5. The personalized review recommendation method based on implicit dimension mining as described in claim 1, characterized in that: The generation of comment embedding and demand embedding includes the following steps: Load the M3E-base-TextDimension model; For each comment text, the comment dimension corresponding to the comment text is directly appended to the end of the comment text to obtain the text in the form of comment and comment dimension; For each user requirement text, the requirement dimensions obtained in step S3 are directly appended to the end of the requirement text to obtain text in the form of requirements and requirement dimensions; Combine all comments and comment dimension texts, as well as requirements and requirement dimension texts, into a list, and input the list of texts one by one into the M3E-base-TextDimension model; The output text embeddings are arranged into a matrix, with the order being consistent with the input text order, and separate comment embeddings and demand embeddings are generated based on the embedding matrix; Save the comment embedding and the request embedding.
6. A personalized review recommendation system based on implicit dimension mining, based on the personalized review recommendation method based on implicit dimension mining as described in any one of claims 1 to 5, characterized in that: It also includes, The fine-tuning module is used to construct the M3E-base-TextDimension model by filtering online product reviews and fine-tuning the M3E-base model. The comment dimensions module is used to identify entities in comments using a large language model and rewrite comments to extract comment dimensions; The Demand Dimension module is used to combine user needs and comment dimensions, and use a large language model to filter out a set of comment dimensions that meet user needs as demand dimensions. The generation module is used to input the comment dimension and demand dimension into the M3E-base-TextDimension model to generate comment embedding and demand embedding; The recommendation module is used to perform recommendations by calculating the cosine similarity between the comment embedding and the demand embedding. Comments and recommendations.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the personalized comment recommendation method based on implicit dimension mining as described in any one of claims 1 to 5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the personalized comment recommendation method based on implicit dimension mining as described in any one of claims 1 to 5.