A small sample review data driven product key user demand mining method

Abstract

The application discloses a small sample review data driven product key user demand mining method, first, a crawler tool is used to obtain product user reviews from an e-commerce website, and an initial user review dataset is established; then, the initial user review dataset is preprocessed to obtain a product user review dataset; next, user review topic clustering is carried out based on an ERNIE model to obtain a plurality of topic user review datasets; then, the topic user review datasets are secondarily processed; finally, key product information phrases are extracted based on an ISIFRank (Improved SIFRank) model to establish a key user demand set. The final result of the application is presented in the form of the key user demand set, and relevant personnel of product design and decision can obtain reference information supporting their own work from the key user demand set.

Description

Technical Field

[0001] This invention belongs to the field of data mining technology, specifically relating to a method for mining key user needs of a product. Background Technology

[0002] In today's fiercely competitive market, new products emerge constantly as consumer demands evolve. To enhance the competitiveness of these new products, companies need to iterate and upgrade their offerings based on user needs. With the development of the internet, more and more users are posting reviews and feedback on products on shopping websites. These reviews generate a large amount of text containing rich, potentially useful information, providing a reliable source for product iteration and upgrades.

[0003] The intervention of user review data transcends the limitations of limited sample statistical prediction and estimation. Through the collection, management, and mining of user reviews, it is possible to effectively quantify users' emotional information and needs, thereby driving iterative product innovation and providing new models and methods for industrial design. From a product perspective, the "product-review data-product" innovative design model uses user review data as the basis for continuous product development, suitable for the iterative upgrades of new products. Hidden within the user review data of new products are users' implicit needs; driven by data, key needs can be uncovered, propelling a new round of product iteration and upgrades. Next-generation products can continuously obtain review feedback data, effectively realizing the value of user review data in a dynamic iterative cycle.

[0004] Existing user demand mining methods focus on common products that have undergone several iterations and have a large number of user reviews available online for uncovering user needs. However, due to the short time new products have been on the market and the small number of buyers, these methods cannot provide a large data source. Therefore, it is necessary to adopt effective methods to analyze small sample review data, uncover key user needs, and provide a reference for the iterative upgrade of new products.

[0005] The key technology for acquiring user needs has gradually shifted from traditional methods to analyzing user comment text using natural language processing (NLP) techniques. The development of NLP technology is essentially about the continuous evolution of natural language models to more accurately represent them. In the process of developing from machine learning to deep learning, some scholars have improved or combined different NLP techniques to learn and express the semantics of user comments more comprehensively, and then introduced them into applied research in the field of user needs mining.

[0006] This invention addresses the problem of user demand mining during the iterative upgrade process of novel products. The number of online user reviews for new products is limited, thus offering limited information for machine learning. Recent advancements in pre-trained language models have demonstrated advantages in this area. By learning human language knowledge in advance from massive datasets and then applying it to specific tasks, valuable information can be extracted from limited user review data. Meanwhile, the application of pre-trained language models in user demand mining is still in the exploratory stage, allowing for new exploratory attempts to assist designers and developers in achieving innovative product iterative design through human-machine collaboration. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides a method for mining key user needs of products driven by small sample review data. First, web crawlers are used to obtain user reviews of products from e-commerce websites, establishing an initial user review dataset. Then, the initial user review dataset is preprocessed to obtain a product user review dataset. Next, user review topics are clustered based on the ERNIE model to obtain several topic user review datasets. These topic user review datasets are then processed again. Finally, key product information phrases are extracted based on the ISIFRank (Improved SIFRank) model to establish a set of key user needs. The final result of this invention is presented as a set of key user needs, from which relevant personnel in product design and decision-making can obtain reference information to support their work.

[0008] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0009] Step 1: Use web crawlers to obtain user reviews of products from e-commerce websites and build an initial user review dataset;

[0010] Step 2: Preprocess the initial user review dataset to obtain a product user review dataset for topic clustering of user reviews;

[0011] Step 2-1: Remove default and duplicate reviews;

[0012] Step 2-2: Remove invalid characters, spaces, and garbled text from user comments to complete the cleaning of user text data;

[0013] Steps 2-3: Identify spelling errors, lexical errors, syntactic errors, and semantic errors in the comments, and complete the error correction;

[0014] Steps 2-4: Merge different descriptions of the same entity and use the same entity name uniformly;

[0015] Step 3: Cluster user comment topics based on the ERNIE model to obtain several topic user comment datasets;

[0016] Step 4: Perform secondary processing on the topic user comment dataset;

[0017] Step 4-1: Use the THULAC tool to segment the user comment text in the topic user comment dataset;

[0018] Step 4-2: Use the NLTK tool to perform part-of-speech tagging on the word segments;

[0019] Step 4-3: Based on Chinese word usage habits and part-of-speech matching principles, eight lexical rules reflecting user needs are set. Regular expressions are used to aggregate these rules to obtain product information phrases, forming a product information phrase set PIP = {PIP1, PIP2, ..., PIP...}. m}; PIP1~PIP m These represent the first to the mth product information phrases, respectively.

[0020] Among them, the eight lexical rules that reflect user needs include: four lexical rules each for adjectives and verbs.

[0021]

[0022] Where JJ represents adjectives, NN represents nouns, RB represents adverbs, VB represents verbs, and DT represents determiners. The symbols are based on the part-of-speech tagging symbols of the NLTK tool.

[0023] Step 5: Extract key product information phrases based on the ISIFRank model and establish a set of key user needs.

[0024] Furthermore, the user comment topic clustering based on the ERNIE model yields several topic user comment datasets, including:

[0025] Step 3-1: Construct the input vector of user comment text, convert the user comment text into a vector, and feed it into the ERNIE model for training and learning;

[0026] The input vector of user comment text is generated by fusing three parts: word embedding, paragraph embedding and position embedding.

[0027] Step 3-2: Based on the online comment encoding and user comment knowledge integration, train the input vector of the user comment text to obtain the word vector representation of the user comment;

[0028] Step 3-3: Connect the Softmax layer to support the topic clustering task of user comments.

[0029] Furthermore, the step of training the input vector of the user comment text based on user online comment encoding and user comment knowledge integration to obtain the word vector representation of the user comment includes:

[0030] Step 3-2-1: Build a pre-trained language model based on the bidirectional Transformer encoder as the basic unit. The Transformer encoder encodes contextual information for each word in the user's comment using a self-attention mechanism.

[0031] Step 3-2-1-1: Through the mapping matrix W Q W K W V This layer transforms the input vector X given by the embedding layer into an input matrix, including query Q, key K, and value V, as shown in the following formula:

[0032] Q = XWQ (1)

[0033] K = XW K (2)

[0034] V = XW V (3)

[0035] Step 3-2-1-2: Calculate the weights by performing similarity calculations on Q and K using the dot product;

[0036] Step 3-2-1-3: Normalize the obtained weights using the softmax function;

[0037] Step 3-2-1-4: The weights and their corresponding values ​​(V) are weighted and summed to obtain the final attention value, i.e., the attention function. The specific formula is as follows:

[0038]

[0039] Step 3-2-1-5: The bidirectional Transformer encoder uses a multi-head mechanism, which completes the calculation of equations (1) to (4) by calling different mapping matrices in parallel, and obtains multiple corresponding attention values, thereby obtaining the unique feature information of each character in the text sequence in different representation subspaces; finally, the attention values ​​are concatenated and linearly mapped to obtain the output y of this layer, and the specific formula is as follows:

[0040]

[0041] y = concat(H1, H2, ..., H h W 0 (6)

[0042] in, Representing different mapping matrices, concat(H1, H2, ..., H... h W represents concatenating h matrices into a single long matrix. 0 Represents the weight matrix;

[0043] Step 3-2-2: Randomly mask characters using a multi-stage knowledge masking strategy, including three masking strategies: character level, entity level, and phrase level. Train the model to predict the masked parts, learn the contextual information of the masked parts, and represent the word vectors of the user's comment.

[0044] Furthermore, the extraction of key product information phrases based on the ISIFRank model to establish a key user needs set includes:

[0045] Step 5-1: Replace ELMo in the SIFRank model with ERNIE and improve it. Obtain word vector representations of topic user comments based on ERNIE.

[0046] Step 5-2: Obtain sentence vector representations of product information phrases and topic user comments based on the SIF algorithm;

[0047] The SIF algorithm calculates a sentence vector by weighting each word vector. The specific formula for the weight of each word vector is as follows:

[0048]

[0049] Where a is a constant, and p(w) is the word frequency;

[0050] The resulting comment text sentence vector v s Represented as:

[0051]

[0052] Similarly, the sentence vector v of the product information phrase is calculated. PIP ;

[0053] Step 5-3: Extract key product information phrases based on the Rank algorithm;

[0054] For a given topic of user reviews of a product, the order of product information phrases depends on the topic user review sentence vector v. s The sentence vector v of product information phrases PIP Cosine similarity between them:

[0055] RANK(v PIPi v s ) = Sim(v PIPi v s (9)

[0056]

[0057] A threshold ε is set, and product information phrases with a cosine similarity greater than ε are extracted as the final key product information phrases.

[0058] Step 5-4: Combining expert knowledge and domain knowledge of specific products, string together scattered key product information phrases and transform them into complete key user requirement statements based on natural language descriptions, thus establishing a complete set of key user requirements;

[0059] For each topic, user comments establish a subset s of key user needs, which consists of three elements: key user need phrase KeyPIP, normalized weight w, and natural language description NLD, i.e., S = {Key PIP, w, NLD}.

[0060] The set of key user requirements A consists of all key user requirement subsets S, i.e., A = {S1, s2, ..., s3}. i}

[0061] Preferably, a is set to [10 -3 10 -4 ].

[0062] The beneficial effects of this invention are as follows:

[0063] This invention, supported by pre-trained language models and driven by small sample review data, mines key user needs to support the iterative upgrades of new products. The final results of this method are presented as a set of key user needs, from which product design and decision-making personnel can obtain reference information to support their work.

[0064] This invention, supported by the ERNIE-ISIFRank model, can easily and efficiently mine user needs from a limited sample of user reviews. ERNIE is used to cluster disorganized user reviews according to different topic tags, its core being the vector representation of user reviews. Unlike other models, this model pre-learns external knowledge, ensuring that the generated user review word vectors contain richer semantics. ISIFRank is used to extract key product information phrases from topic-based user reviews, thus supporting the establishment of a key user need set. The improvement of ISIFRank lies in its integration with ERNIE, inheriting its advantages and establishing new keyword extraction rules suitable for the product design domain, ensuring the extraction of semantically more complete key product information phrases, thereby supporting the establishment of a key user need set. Attached Figure Description

[0065] Figure 1 This is a flowchart of the method of the present invention.

[0066] Figure 2 This embodiment of the invention is based on the ERNIE user comment topic clustering model.

[0067] Figure 3 This is a structural diagram of the Transformer encoding unit in an embodiment of the present invention.

[0068] Figure 4 This is a multi-stage knowledge masking strategy according to an embodiment of the present invention.

[0069] Figure 5 This embodiment of the invention establishes a model based on the key user requirement set of ISIFRank.

[0070] Figure 6 This is the set of key user requirements in the embodiments of the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0072] A method for mining key user needs of a product driven by small sample review data includes the following steps:

[0073] Step 1: Use web crawlers to obtain user reviews of products from e-commerce websites and build an initial user review dataset;

[0074] Step 2: Preprocess the initial user review dataset to obtain a product user review dataset for topic clustering of user reviews;

[0075] Step 2-1: Remove useless duplicate data such as default reviews, duplicate reviews, and fake reviews, and keep valuable comments;

[0076] Step 2-2: Remove invalid characters, spaces, and garbled text from user comments to clean the user text data;

[0077] Steps 2-3: Identify spelling errors, lexical errors, syntactic errors, and semantic errors in the comments, and complete the error correction;

[0078] Steps 2-4: Merge different descriptions of the same entity and use the same entity name uniformly;

[0079] After the above series of preprocessing steps, a relatively high-quality new product user review dataset for user review topic clustering is obtained.

[0080] Step 3: Cluster user comment topics based on the ERNIE model to obtain several topic user comment datasets;

[0081] Step 3-1: Construct the input vector of the user comment text, transform the user comment text into a vector that the computer can understand, and feed it into the model for training and learning. The input vector of the user comment text includes three parts: word embedding, segment embedding and position embedding, which are fused to generate the input vector of the user comment text.

[0082] Step 3-2: Based on the online comment encoding and user comment knowledge integration, train the input vector of the user comment text to obtain the word vector representation of the user comment;

[0083] Step 3-2-1: Build a pre-trained language model based on the bidirectional Transformer encoder as the basic unit. The Transformer encoder encodes contextual information for each word in the user's comment using a self-attention mechanism.

[0084] Step 3-2-1-1: Through the mapping matrix W Q W K W V This layer transforms the input vector X given by the embedding layer into an input matrix, including query Q, key K, and value V, as shown in the following formula:

[0085] Q = XW Q (1)

[0086] K = XW K (2)

[0087] V = XW V (3)

[0088] Step 3-2-1-2: Calculate the weights by performing similarity calculations on Q and K using the dot product;

[0089] Step 3-2-1-3: Normalize the obtained weights using the softmax function;

[0090] Step 3-2-1-4: The weights and their corresponding values ​​(V) are weighted and summed to obtain the final attention value, i.e., the attention function. The specific formula is as follows:

[0091]

[0092] Step 3-2-1-5: The bidirectional Transformer encoder uses a multi-head mechanism, which completes the calculation of equations (1) to (4) by calling different mapping matrices in parallel, and obtains multiple corresponding attention values, thereby obtaining the unique feature information of each character in the text sequence in different representation subspaces; finally, the attention values ​​are concatenated and linearly mapped to obtain the output y of this layer, and the specific formula is as follows:

[0093]

[0094] y = concat(H1, H2, ..., H h W 0 (6)

[0095] in, Representing different mapping matrices, concat(H1, H2, ..., H... h W represents concatenating h matrices into a single long matrix. 0 Represents the weight matrix;

[0096] Step 3-2-2: Randomly mask characters using a multi-stage knowledge masking strategy, including three masking strategies: character level, entity level, and phrase level. Train the model to predict the masked parts, learn the contextual information of the masked parts, and represent the word vectors of the user's comment.

[0097] Step 3-3: Connect the softmax layer to support the topic clustering task of user comments.

[0098] Step 4: Perform secondary processing on the topic user comment dataset;

[0099] Step 4-1: Use the THULAC tool to segment the user comment text in the topic user comment dataset;

[0100] Step 4-2: Use the NLTK tool to perform part-of-speech tagging on the word segments;

[0101] Step 4-3: Based on everyday language used in product design and according to Chinese vocabulary usage habits and part-of-speech matching principles, eight lexical rules reflecting user needs are established. Regular expressions are then used to aggregate these rules to obtain product information phrases, forming a product information phrase set PIP = {PIP1, PIP2, ..., PIP...}. m};

[0102] Among them, the eight lexical rules that reflect user needs include: four lexical rules each for adjectives and verbs.

[0103]

[0104] Where JJ represents adjectives, NN represents nouns, RB represents adverbs, VB represents verbs, and DT represents determiners. The symbols are based on the part-of-speech tagging symbols of the NLTK tool.

[0105] Step 5: Extract key product information phrases based on the ISIFRank model and establish a set of key user needs.

[0106] Step 5-1: Replace ELMo in the SIFRank model with ERNIE and improve it. Obtain word vector representations of topic user comments based on ERNIE.

[0107] Step 5-2: Obtain sentence vector representations of product information phrases and topic user comments based on the SIF algorithm;

[0108] The SIF algorithm calculates a sentence vector by weighting each word vector. The specific formula for the weight of each word vector is as follows:

[0109]

[0110] Where a is a hyperparameter and p(w) is the word frequency;

[0111] The resulting comment text sentence vector v s Represented as:

[0112]

[0113] Similarly, the sentence vector v of the product information phrase is calculated. PIP ;

[0114] Step 5-3: Extract key product information phrases based on the Rank algorithm;

[0115] For a given topic of user reviews of a product, the order of product information phrases depends on the topic user review sentence vector v. s The sentence vector v of product information phrases PIP Cosine similarity between them:

[0116] RANK(v PIPi v s ) = Sim(v PIPi v s (9)

[0117]

[0118] A threshold ε is set, and product information phrases with a cosine similarity greater than ε are extracted as the final key product information phrases.

[0119] Step 5-4: Combining expert knowledge and domain knowledge of specific products, string together scattered key product information phrases and transform them into complete key user requirement statements based on natural language descriptions, establishing a complete set of key user requirements that designers can refer to.

[0120] For each topic, a subset S of key user needs is built from user comments. It consists of three elements: key user need phrase KeyPIP, normalized weight w, and natural language description NLD, i.e., S = {Key PIP, w, NLD}.

[0121] The set of key user requirements A consists of all subsets S of key user requirements, i.e., A = {S1, S2, ..., S...} i} Specific implementation examples:

[0123] Reference Figures 1-6 The specific steps of this invention's novel product key user needs mining method driven by small sample review data are as follows:

[0124] In recent years, with the increasing number of cat owners in China, the pet-related industry has developed rapidly, and products specifically designed for cats have become more diversified. Smart cat feeders are a popular new product in the past year. This paper uses a smart cat feeder as an example to verify the technical implementation and application of the ERNIE-ISIFRank model.

[0125] I. Preparation of the initial user review dataset

[0126] Adhering to the principles of both high authority and excellent review quality, authoritative e-commerce websites were selected as the source of online reviews, primarily including JD.com, Suning.com, and Tmall. Web scraping tools were used to collect online user reviews of the Homan Real smart cat feeder, completing the initial user review dataset construction. This product was launched in November 2021, and the user review collection period spanned from December 15, 2021 to June 31, 2022, containing a total of 6754 user reviews.

[0127] II. User Comment Text Preprocessing

[0128] The process involves removing useless duplicate data such as default reviews, duplicate reviews, and fake reviews, leaving only valuable comments; cleaning user text data by removing invalid characters, spaces, and garbled text from user comments; identifying and correcting spelling errors, lexical errors, syntactic errors, and semantic errors in the comments; and merging different descriptions of the same entity and using the same entity name. After this series of preprocessing steps, a new type of product user review dataset is obtained.

[0129] Based on the characteristics of smart cat feeders, and after professional discussion, review categorization tags were determined from three dimensions: appearance, functionality, and emotional needs. Eight tags, ranging from 0 to 7, were assigned, representing shape, color, material, interactive operation, smart feeding, installation and maintenance, cost-effectiveness, and practicality, respectively. The new product user review dataset contains 5000 manually labeled user reviews, which were randomly divided into training and testing sets at an 8:2 ratio. The product's user review tags and examples are shown in Table 1.

[0130] Table 1 User Comment Tags and Examples

[0131]

[0132] III. User Comment Topic Clustering Based on ERNIE Model

[0133] Given an input sequence of user comment text T = (t1, t2, ..., t3) n ), can t n Treating the input text as a single Chinese character, for example, T represents the input text "Simple design, I like it very much.", we insert [CLS] and [SEP] symbols into the input text, converting it into word embeddings and segment embeddings, and then combining them with positional embeddings. Since all three have the same dimension, they are summed through the ERNIE input layer to obtain the input vector X = (x1, x2, ..., x...). n ).

[0134] ERNIE consists of 12 encoder layers, 768 hidden units, and 12 attention heads. Each layer includes computational mechanisms such as summation, normalization, feedforward neural networks, and self-attention. Figure 3 As shown.

[0135] A pre-trained language model is built using a bidirectional Transformer encoder as the basic unit. The Transformer encoder relies on a self-attention mechanism to encode contextual information for each word in the user's comment. This is achieved through a mapping matrix W. Q W K W V The input vector X given by the embedding layer is transformed into an input matrix: query Q, key K, and value V, where Q = XW. Q K = XW K V = XW V The weights are calculated by performing similarity calculations on Q and K using the dot product. These weights are then normalized using the softmax function. Finally, the weights are summed with their corresponding values ​​in V to obtain the final attention value.

[0136] The encoder employs a "multi-head" mechanism, performing the above calculations by calling different mapping matrices in parallel to obtain multiple corresponding attention values. This allows it to acquire unique feature information of each character in the text sequence across different representation subspaces. Finally, the attention values ​​are concatenated and linearly mapped to obtain the output y of this layer. y = concat(H1, H2, ..., H h W 0 .

[0137] By inputting X into ERNIE and combining it with a multi-stage knowledge masking strategy to randomly mask characters (including character-level, entity-level, and phrase-level masking strategies), the model is trained to predict the masked parts, learn the contextual information of the masked parts, and represent the word vectors of user comments, such as... Figure 4 As shown, after the ERNIE main structure is pre-trained on its own pre-training corpus for multiple tasks, the state H = (h1, h, ..., h) of the last Transformer layer is taken. n The word vector representations of user comments are used. These word vector representations are then input into a softmax layer for topic clustering of user comments. The clustering results are saved in eight txt files according to the label order, resulting in eight topic-based user comment datasets.

[0138] IV. Secondary Processing of User Comment Dataset

[0139] Taking the topic of "appearance" as an example, the topic user comment dataset is processed in a second way. The THULAC tool is used for word segmentation, and the NLTK tool is used for part-of-speech tagging. Combined with regular expressions and the eight new product comment part-of-speech aggregation rules proposed in this invention, the word segmentation results are aggregated to obtain product information phrases (PIP) examples as shown in Table 2.

[0140] Table 2 Examples of secondary processing of topic user comments

[0141]

[0142] V. Extraction of Key Product Information Phrases Based on the ISIFRank Model

[0143] The word vector representation of user comments on the topic of "appearance" was obtained based on the ERNIE algorithm.

[0144] The SIF algorithm is used to obtain sentence vector representations of product information phrases and topic user reviews. The SIF algorithm calculates a weighted average of each word vector to obtain the sentence vector. The formula for the weight of each word vector is: The resulting sentence vector v of the comment text obtained through this weighting method s Represented as The same method can be used to calculate the sentence vector v of the product information phrase. PIP .

[0145] Calculate the vector v of user comments on the topic s The sentence vector v of product information phrases PIP Cosine similarity between them

[0146] Set a threshold ε = 0.45, extract product information phrases with a cosine similarity greater than ε, return the key product information phrases in the "Appearance" topic in descending order, and manually remove incorrectly extracted product information phrases.

[0147] All the topic user review datasets were processed using the above method to extract key product information phrases (Key PIPs), and the final results are shown in Table 3.

[0148] Table 3 Results of Key Product Information Phrases Extraction

[0149]

[0150] VI. Establish a set of key user requirements

[0151] The extracted key product information phrases represent product features that users pay close attention to, and are important data for analyzing and building a set of key user needs.

[0152] A comprehensive analysis of key product information phrases within each theme, such as those concerning design, color, material, and installation / maintenance, reveals that all key product information phrases contain only positive descriptive words. This indicates that users are generally satisfied, suggesting that existing products largely meet user needs and thus represent satisfied needs. However, in areas such as interactive operation, intelligent feeding, cost-effectiveness, and practicality, key product information phrases contain both positive and negative descriptive words. This indicates that user product experiences are inconsistent, suggesting that existing products do not fully meet the needs of most users and that there is still significant room for improvement, representing unmet needs.

[0153] Satisfied and unsatisfied needs together constitute all key user requirements. Combining relevant product knowledge of smart feeders, the scattered key product information phrases within each topic are summarized and merged, then linked and described in natural language, transforming them into complete and easily understandable key user requirements. To further clarify the presentation, the weights of the key user requirement phrases within each topic are normalized, establishing a subset of key user requirements.

[0154] The key user requirements subset for each topic was constructed using the method described above, and the union of all subsets constituted the final key user requirements set, as shown below. Figure 6 As shown.

[0155] In specific examples of new product applications, comparative experiments can demonstrate the effectiveness of the present invention.

[0156] For the ERNIE-based topic user demand clustering model, SVM, BiLSTM, and BERT were selected as baseline algorithms; for the ISIFRank-based key user demand set model, TF-IDF, TextRank, and SIFRANK were selected as baseline algorithms. Precision, recall, and F1 score were used to evaluate the model performance. Ten-fold cross-validation was used, and the results were averaged after 10 runs, as shown in Figure 4.

[0157] Table 4 Model Evaluation Indicators

[0158]

[0159] In user comment topic clustering algorithms, ERNIE achieved the highest accuracy (0.9248) and recall (0.8845). SVM and BiLSTM rely heavily on word vector quality and lack generalization ability, resulting in significantly lower performance than pre-trained language models BERT and ERNIE. Both SVM and BiLSTM, based on the Transformer bidirectional encoder mechanism, can adaptively extract features from user comment text, improving classification accuracy. Among pre-trained language models, ERNIE outperforms BERT because its masking mechanism is richer, allowing it to learn deeper semantic information, especially with limited data.

[0160] In key product information extraction algorithms, ISIFRank achieved the highest accuracy (0.9049) and recall (0.8143). TF-IDF and TextRank performed worse than SIFRank because topic words do not appear consistently in a user review, and these two algorithms, being frequency-based keyword extraction algorithms, cannot understand the semantics of the review text, leading to inaccurate extraction of key product information phrases. Considering this limitation, SIFRank uses the pre-trained language model ElMo to acquire various grammatical and semantic information of words, thereby ensuring the extraction of high-quality key product information phrases. Furthermore, SIFRank also compensates for the shortcomings of TF-IDF and TextRank. TextRank has the drawback of failing to aggregate two words together to fully express key product information phrases. The ISIFRank proposed in this invention is superior to SIFRank mainly due to the following two improvements: First, by introducing the pre-trained model ERNIE to replace ELMo, it improves the poor inference efficiency when acquiring user review word vectors. ELMo is based on BiLSTM, and comparative experiments have confirmed that ERNIE outperforms BiLSTM, as shown in Table 1. Second, this invention, tailored to the linguistic characteristics of product reviews, establishes new part-of-speech aggregation rules around verbs and adjectives, thus further ensuring the completeness and accuracy of the extracted key product information phrases.

[0161] The F1 score is an evaluation metric that considers both precision and recall, comprehensively reflecting the overall performance of an algorithm. ERNIE and ISIFRank achieved F1 scores of 0.8934 and 0.8534, respectively, both reaching the highest values. Combining these two top-performing algorithms, and using ERNIE-ISIFRank for key user need mining, ensures accurate and efficient establishment of key user need sets. This invention also broadens the feasible field methods for mining user needs for new products based on pre-trained language models.

Claims

1. A method for mining key user needs of a product driven by small sample review data, characterized in that, Includes the following steps: Step 1: Use web crawlers to obtain user reviews of products from e-commerce websites and build an initial user review dataset; Step 2: Preprocess the initial user review dataset to obtain a product user review dataset for topic clustering of user reviews; Step 2-1: Remove default and duplicate reviews; Step 2-2: Remove invalid characters, spaces, and garbled text from user comments to complete the cleaning of user text data; Steps 2-3: Identify spelling errors, lexical errors, syntactic errors, and semantic errors in the comments, and complete the error correction; Steps 2-4: Merge different descriptions of the same entity and use the same entity name uniformly; Step 3: Perform topic clustering of user comments based on the ERNIE model to obtain several topic-based user comment datasets, including: Step 3-1: Construct the input vector of user comment text, convert the user comment text into a vector, and feed it into the ERNIE model for training and learning; The input vector of user comment text is generated by fusing three parts: word embedding, paragraph embedding and position embedding. Step 3-2: Train the input vector of user comment text based on user online comment encoding and user comment knowledge integration to obtain the word vector representation of user comments; Step 3-3: Connect the Softmax layer to support the topic clustering task of user comments; Step 4: Perform secondary processing on the topic user comment dataset; Step 4-1: Use the THULAC tool to segment the user comment text in the topic user comment dataset; Step 4-2: Use the NLTK tool to perform part-of-speech tagging on the word segments; Step 4-3: Based on Chinese word usage habits and part-of-speech matching principles, eight lexical rules reflecting user needs are established. Regular expressions are then used to aggregate these rules to obtain product information phrases, forming a set of product information phrases. ; PIP 1 ~PIP m They represent the 1st to the 1st. m A short product information phrase; Among them, the eight lexical rules that reflect user needs include: four lexical rules each for adjectives and verbs. Where JJ represents adjectives, NN represents nouns, RB represents adverbs, VB represents verbs, and DT represents determiners. The symbols are based on the part-of-speech tagging symbols of the NLTK tool. Step 5: Extract key product information phrases based on the ISIFRank model and establish a key user needs set, including: Step 5-1: Replace ELMo in the SIFRank model with ERNIE and improve it. Obtain word vector representations of topic user comments based on ERNIE. Step 5-2: Obtain sentence vector representations of product information phrases and topic user comments based on the SIF algorithm; The SIF algorithm calculates a sentence vector by weighting each word vector. The specific formula for the weight of each word vector is as follows: (7) in It is a constant. For word frequency; The resulting comment text sentence vector Represented as: (8) Similarly, the sentence vectors of the product information phrases are calculated. ; Step 5-3: Extract key product information phrases based on the Rank algorithm; For a given topic of user reviews, the order of product information phrases depends on the topic user review sentence vector. Sentence vectors of product information phrases Cosine similarity between them: (9) (10) Set threshold The cosine similarity is greater than Extract product information phrases and use them as the final key product information phrases; Step 5-4: Combining expert knowledge and domain knowledge of specific products, string together scattered key product information phrases and transform them into complete key user requirement statements based on natural language descriptions, thus establishing a complete set of key user requirements; For each topic, user reviews establish a subset of key user needs. Composed of key user requirement phrases Normalized weights and natural language description It consists of three elements, namely ; Key User Requirements Set From all key user needs subset Composition, that is .

2. The method for mining key user needs of a product driven by small sample review data according to claim 1, characterized in that, The method of training the input vector of user comment text based on user online comment encoding and user comment knowledge integration to obtain the word vector representation of user comments includes: Step 3-2-1: Construct a pre-trained language model based on the bidirectional Transformer encoder as the basic unit. The Transformer encoder encodes contextual information for each word in the user's comment using a self-attention mechanism. Step 3-2-1-1: Through the mapping matrix , , This layer takes the input vector given by the embedding layer. Transform each into an input matrix, including the query. ,key Sum The specific formula is as follows: (1) (2) (3) Step 3-2-1-2: Through dot product pairs and The weights are obtained by performing similarity calculations; Step 3-2-1-3: Normalize the obtained weights using the softmax function; Step 3-2-1-4: Combine the weights and corresponding... The final attention value, or attention function, is obtained by weighted summation, and the specific formula is as follows: (4) Step 3-2-1-5: The bidirectional Transformer encoder adopts a multi-head mechanism, which completes the calculation of equations (1) to (4) by calling different mapping matrices in parallel, and obtains multiple corresponding attention values, thereby obtaining the unique feature information of each character in the text sequence in different representation subspaces; finally, the attention values ​​are concatenated and linearly mapped to obtain the output of this layer. The specific formula is as follows: (5) (6) in, , , Representing different mapping matrices, Indicates will The matrices are concatenated to form a long matrix. Represents the weight matrix; Step 3-2-2: Randomly mask characters using a multi-stage knowledge masking strategy, including three masking strategies: character level, entity level, and phrase level. Train the model to predict the masked parts, learn the contextual information of the masked parts, and represent the word vectors of the user's comment.

3. The method for mining key user needs of a product driven by small sample review data according to claim 1, characterized in that, The Set as .

Images