Large-model-based open question point mining method for user comments

Abstract

The invention discloses a large-model-based open question point mining method for user comments, which comprises the following steps: S1, user comment input and preprocessing: acquiring an original user comment data source, and preprocessing the original user comment data source to obtain effective original user comments; s2, performing large-model fine-grained sentiment analysis; s3, generating an open question point abstract; s4, encoding question point sentence vectors; s7, performing label reinjection and association; and S8, carrying out tag statistical analysis. The method has the advantages that the whole process is automatic, model training and data labeling are not needed, cost is low, universality is high, problem mining is accurate and comprehensive, implicit and emerging problems are covered, labels are standardized, noise and fragmentation are avoided, mining results can be quantified and traceable, accurate decision is supported, and the method is suitable for multi-field scenes.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically relating to an open-ended question point mining method based on a large model for user comments. Background Technology

[0002] In today's digital age, user reviews have become a core data source for businesses to understand market feedback and optimize products and services. They also serve as a crucial basis for product and service improvement, making them vital for business decision-making. Traditional user review analysis methods often rely on manual annotation and simple text classification techniques, which are not only inefficient but also fail to comprehensively and deeply uncover potential problems within user reviews. With the development of deep learning and large-scale modeling technologies, new possibilities have emerged for automated and intelligent processing of user reviews. However, currently, there is no complete and efficient open-ended problem-point mining solution based on large-scale models that can accurately extract valuable problem information from user reviews without involving complex model training and data annotation.

[0003] Existing user comment analysis methods have the following limitations: (1) Reliance on a pre-defined tag system: It cannot automatically discover emerging problems that are not pre-defined (such as sudden quality problems or scenario-based pain points). Traditional user comment analysis often adopts the "pre-defined tags and text classification" model, that is, a fixed set of tags is first defined manually, and then the classification model is trained to match the comments to the corresponding tags. The core limitation of this model is "tag first". If emerging problems that are not pre-defined appear in user feedback, the model cannot identify and classify them. It can only rely on manual addition of tags and retraining, which leads to a delay in the discovery of emerging problems and misses the best opportunity for product improvement or crisis prevention. (2) High cost of manual labeling: Traditional problem classification requires manual definition of labels and labeling of training data, which is difficult to cover long-tail problems. Traditional methods require manual labeling of a large amount of training data to train an effective model, which is extremely costly. According to industry standards, each comment needs 2-3 minutes to label, and 10,000 comments need 333-500 man-hours. If multiple fields of products are involved, different scenario data need to be labeled repeatedly, which further increases the cost. (3) Insufficient depth of problem mining: Traditional methods are difficult to accurately locate high-value negative issues from massive comments, especially the ability to identify implicit complaints is weak. Traditional methods rely on keyword matching (such as classifying comments containing negative words such as "bad", "inferior", "not good" as problem comments) or simple sentiment analysis, which cannot identify "implicit complaints", that is, comments in which users do not use explicit negative words but actually express dissatisfaction; (4) Label redundancy and noise: Directly generated labels may contain a large number of non-problem labels (such as functional descriptions and neutral evaluations), which dilutes the key issues. Traditional models directly generate labels from all comments without “problem screening”, resulting in a large number of non-problem labels being mixed in. (5) High dependence on human intervention: The labels after clustering need to be manually summarized and named, which is inefficient and highly subjective. The traditional method of "clustering and label naming" requires a lot of human intervention. That is, the comments are first clustered by algorithms such as K-means, and then the labels are manually named for each cluster. In this process, the subjectivity of manual naming is strong and the efficiency is low. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a fully automated, open-ended problem point mining method for user reviews based on a large model, which requires no model training or data annotation.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: An open-ended question point mining method based on a large model for user comments includes the following steps: S1. User comment input and preprocessing: Obtain the original user comment data source and preprocess the original user comment data source to obtain valid original user comments; S2. Fine-grained sentiment analysis of large model: Call the pre-trained large model to perform sentiment analysis on each of the valid original user comments obtained in step S1, and output the sentiment polarity and sentiment dimension of each valid original user comment. The sentiment polarity includes negative and non-negative. S3. Open-ended question summary generation: Select valid original user comments with negative sentiment polarity from step S2 as target comments, and call the pre-trained large model to generate question summary for each target comment through prompting engineering; S4. Question point sentence vector encoding: Use a pre-trained sentence vector model to convert all the question point summaries generated in step S3 into semantic vectors; S5. Problem point clustering and noise reduction: The DBSCAN clustering algorithm is used to cluster the semantic vectors obtained in step S4, merge semantically similar problem point summaries to form clusters, and automatically remove noise points; S6. Automated generation of cluster labels: The pre-trained large model is called to process the problem point summaries in each cluster obtained in step S5 and generate standardized general labels. S7. Tag Back-injection and Association: Establish the mapping relationship between the standardized general tags generated in step S6 and the valid original user comments in step S1, forming a tagged user comment dataset; S8. Tag Statistical Analysis: Based on the tagged user review dataset from step S7, perform tag frequency statistics, time trend analysis, and problem source tracing query.

[0006] Preferably, the original user review data source in step S1 includes at least one of e-commerce platform reviews, social media texts, and customer service conversation records. The preprocessing includes removing duplicate reviews, filtering meaningless text, and assigning a unique review ID to each valid original user review.

[0007] Preferably, in step S2, non-negative includes positive and neutral, and the emotional dimension includes at least one of product quality, service attitude and user experience. In step S2, the pre-trained large model is GPT-4 or Claude 3.5.

[0008] Preferably, the problem summary in step S3 is a core complaint description of no more than 10 words, and does not contain sentimental words or irrelevant details.

[0009] Preferably, the problem tracing query in step S8 involves retrieving all valid original user comments corresponding to a standardized general tag by inputting the standardized general tag.

[0010] Preferably, the pre-trained sentence vector model in step S4 is either Sentence-BERT or Connan-embedding.

[0011] By adopting the above technical solution, the present invention has the following beneficial effects: (1) From step S1 to step S8, all steps of this invention are automatically completed by a large model and algorithm. There is no need for manual annotation of training data, manual screening of question comments, and manual naming of cluster labels. The whole process is automated and no manual intervention is required. Compared with traditional methods, the efficiency is greatly improved, while avoiding subjective errors caused by manual intervention, such as differences in the naming of labels by different annotators. (2) This invention can be used based on pre-trained large model and pre-trained sentence vector model. There is no need to train a special model for a specific domain or label domain-specific data. It is low-cost and highly versatile. (3) For “implicit complaints”, the large model can identify problems without explicit negative words through semantic understanding. For “emerging problems”, the DBSCAN algorithm does not need to preset the number of clusters and can automatically discover emerging problems that are not preset. For “long-tail problems”, even if a certain type of problem only appears a few times, as long as the semantics are similar, it can be clustered into an effective cluster, avoiding the omission of long-tail problems. Problem mining is accurate and comprehensive, covering implicit and emerging problems. (4) This application standardizes the label, eliminating noise and fragmentation; (5) This application is applicable to multiple fields and scenarios, and the mining results are quantifiable and traceable, supporting accurate decision-making; In summary, this invention has the advantages of full-process automation, no need for model training and data annotation, low cost, strong versatility, accurate and comprehensive problem mining, covering implicit and emerging problems, standardized labels, no noise and fragmentation, quantifiable and traceable mining results, support for accurate decision-making, and adaptability to multiple domain scenarios. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0013] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0014] The components of the embodiments of the invention described and shown in the accompanying drawings can typically be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention.

[0015] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0017] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0018] Example 1 Existing user comment analysis methods suffer from limitations such as "limited preset tags, high manual costs, insufficient mining depth, excessive tag noise, and low efficiency," making them ill-suited for the analysis of massive, dynamic, and complex user comments. With the development of pre-trained large-scale models like GPT-4.1 and Claude 3.5, their powerful natural language understanding and generation capabilities offer a new path to solving these problems, enabling fine-grained text analysis and standardized generation without manual annotation or model training. However, currently, there is no complete technical solution that combines large-scale models with clustering algorithms to form a fully automated closed loop of "sentiment screening - issue mining - clustering noise reduction - tag generation - statistical analysis" to address the pain points of traditional methods. Therefore, this invention proposes an open-ended issue mining method for user comments based on a large-scale model, achieving zero-manual intervention, high efficiency, and high accuracy in issue mining. In this embodiment, a fully automated, model-train-free, and open-ended method for mining user issues is proposed. Specific objectives include: 1. Automatically focusing on high-value negative issues and filtering out non-issue information; 2. Achieving unsupervised aggregation and standardized tag generation of semantically similar issues, eliminating human intervention; 3. Establishing a strong correlation between tags and original comments to support quantitative statistics and issue tracing; 4. Adapting to user comment analysis in different fields (consumer electronics, automobiles, government affairs, etc.) and possessing strong versatility.

[0019] like Figure 1 As shown, in one embodiment of the present invention, the open-ended question point mining method based on a large model for user comments includes the following steps: S1. User comment input and preprocessing: Obtain the original user comment data source and preprocess the original user comment data source to obtain valid original user comments. The original user comment data source in step S1 includes at least one of e-commerce platform comments, social media texts and customer service dialogue records. The preprocessing includes removing duplicate comments, filtering meaningless text, and assigning a unique comment ID to each valid original user comment. S2. Fine-grained sentiment analysis of large model: Call the pre-trained large model to perform sentiment analysis on each valid original user comment obtained in step S1, and output the sentiment polarity and sentiment dimension of each valid original user comment. The sentiment polarity includes negative and non-negative. In step S2, non-negative includes positive and neutral. The sentiment dimension includes at least one of product quality, service attitude and user experience. The pre-trained large model in S2 is GPT-4 or Claude 3.5. Specifically, the API structure of the pre-trained large model is called to perform fine-grained sentiment analysis on each original comment after preprocessing in step S1. The core objective is to filter out negative comments (i.e. comments containing user complaints and feedback) and label their sentiment dimensions (for reference in subsequent question classification). More specifically, a large model API can be invoked via an HTTP request, passing in parameters including prompt words and the original comment text. Then, explicit and structured prompt words are used to ensure a consistent output format from the large model. An example is: "Task: Analyze the sentiment tendency and sentiment dimension of user comments. Sentiment tendency outputs only 'negative' or 'non-negative' (non-negative includes positive and neutral); sentiment dimension selects only the most matching category from 'product quality,' 'service attitude,' 'user experience,' and 'logistics and after-sales service.'" The results returned by the large model are then parsed, filtering comments with a "negative" sentiment polarity as "target comments," retaining their comment ID, original text, and sentiment dimension. Comments with a "non-negative" sentiment polarity are temporarily stored and not included in subsequent question mining (to avoid interference from non-question information). This step efficiently filters "non-question comments" (such as positive reviews and neutral descriptions), focusing on the user's real pain points and solving the problem of "numerous label noise" in traditional methods.

[0020] S3. Open-ended problem summary generation: Select valid original user comments with negative sentiment polarity from step S2 as target comments, call the pre-trained large model to generate problem summary for each target comment through prompting engineering. The problem summary in step S3 is a core complaint problem description of no more than 10 characters, and does not contain sentiment words and irrelevant details. Specifically, for the target comments (negative comments) selected in step S2, the pre-trained large model is invoked again to generate "problem point summaries" through prompting engineering. This involves extracting the core complaint content of each negative comment, removing redundant information (such as sentiment descriptions and irrelevant details), and forming a concise and standardized problem description text. The specific operations include prompt word design, summary optimization, and result output. The core requirements for prompt word design are "focusing on the problem and keeping the length concise" to avoid the large model generating sentiment words or irrelevant descriptions. Summary optimization means that if the output of the large model exceeds 10 characters or contains non-problem information, it is corrected through secondary prompts (such as "Please shorten the above problem description to within 10 characters, keeping only the core problem"). The result output is to form a "problem point summary set", with each summary associated with the ID of the target comment. This step uses the semantic understanding capability of the large model to automatically extract the core problem, solving the problem of "weak implicit complaint recognition" in traditional methods. Even if the comment does not have explicit negative words, the large model can still locate the problem. In this step, the problem point summary is a core complaint problem description of no more than 10 characters.

[0021] S4. Question point sentence vector encoding: All question point summaries generated in step S3 are converted into semantic vectors using a pre-trained sentence vector model. The pre-trained sentence vector model in step S4 can be either Sentence-BERT or Connan-embedding. Specifically, in order to achieve "aggregation of semantically similar questions", the "question point summary" (text form) generated in step S3 needs to be converted into a computer-computable "semantic vector" (numerical form) - that is, by using a pre-trained sentence vector model, each summary is mapped to a high-dimensional vector, and the similarity of the vector corresponds to the similarity of the text semantics. The specific operations are sentence vector model selection, vector generation and vector standardization. The sentence vector model is a mature pre-trained sentence vector model, including but not limited to Sentence-BERT (suitable for general scenarios) and Connan-embedding (suitable for Chinese comment scenarios). The output vector dimension is selected as 512-dimensional or 768-dimensional (the higher the dimension, the stronger the semantic discriminative power, but the computational cost is slightly higher). Vector generation involves inputting each summary from the "problem point summary set" into a sentence vector model, and the model outputs the corresponding semantic vector. Vector normalization is a process of normalizing the generated semantic vectors (adjusting the vector's magnitude to 1) to avoid the impact of differences in vector magnitude on subsequent clustering results, ensuring that clustering is based solely on semantic similarity. The core function of this step is to "transform textual semantics into numerical features," providing a computational foundation for subsequent clustering and addressing the shortcoming of traditional methods that "cannot automatically aggregate semantically similar features."

[0022] S5. Problem point clustering and noise reduction: The DBSCAN clustering algorithm is used to cluster the semantic vectors obtained in step S4, merge semantically similar problem point summaries to form clusters, and automatically remove noise points; Specifically, DBSCAN (density-based spatial clustering algorithm for noise application) is used to cluster the "semantic vectors" obtained in step S4. The core objectives are: 1. to aggregate semantically similar question point summaries into "clusters"; 2. to automatically remove "noise points" (i.e., isolated summaries with no similar questions, appearing only once without other similar comments, and judged as noise). The specific steps are DBSCAN parameter settings, specific clustering process, and noise reduction processing: The DBSCAN parameter settings are adjusted based on the semantic complexity and number of question point summaries. The core parameters include: Neighborhood radius ε: Defines the threshold for "similarity". If the distance between two vectors is less than ε, they are considered "semantically similar". The value range is 0.4-0.6 (ε=0.5 is recommended for Chinese comment scenarios). Minimum sample number minPts: Defines the minimum size of a "cluster". If the ε-neighborhood of a vector contains at least minPts other vectors, a cluster is formed. The value range is 5 - 8 (when the number of comments is greater than 1000, it is recommended that minPts = 5; when the number of comments is less than 1000, it is recommended that minPts = 3). The specific clustering process is to traverse all semantic vectors. For each unlabeled vector, calculate the number of samples in its ε-neighborhood. If the number of samples ≥ minPts, regard this vector as a "core point", and classify all samples in its neighborhood into the same cluster, and then recursively expand this cluster (also classify the neighborhood samples of the samples in the core point's neighborhood into this cluster). If the number of samples < minPts, mark it as a "border point" or "noise point". The noise reduction process is to eliminate the clusters with the number of samples < minPts (i.e., "noise clusters"), and retain the effective clusters with the number of samples ≥ minPts. Each effective cluster corresponds to a type of "semantically similar problems". The advantage of the DBSCAN algorithm is that it "does not require presetting the number of clusters" and can automatically discover the number of problem categories that are unknown (such as the problem of "touch failure after system upgrade" that suddenly appears, which will automatically form a new cluster), solving the defect of the traditional K-means algorithm that "requires presetting the number of clusters and cannot discover emerging problems". At the same time, the noise reduction process can eliminate isolated low-frequency noise problems, ensuring that the subsequent labels focus on high-value problems.

[0023] S6. Automatic generation of clustering labels: Call the pre-trained large model to process the problem point summaries in each clustering cluster obtained in step S5 to generate standardized general labels. Specifically, for each "effective clustering cluster" obtained in step S5, call the pre-trained large model to generate a "standardized general label" - that is, summarize the commonalities of all problem point summaries in this cluster to form a concise and unified label for subsequent statistical analysis and problem location. The specific operation steps include prompt design, label verification, and result output. The core requirement of prompt design is to "summarize commonalities, be concise and unified". An example of a prompt can be "Task: Based on multiple similar descriptions of the following same type of problems, generate 1 general label. Requirements: 1. No more than 6 characters; 2. Contain the core of the problem; 3. The label format is 'XX problem' or 'XX failure' (in the government affairs scenario, it can be changed to 'XX difficult problem'); 4. Can cover the commonalities of all descriptions. The problem description list: [all problem point summaries in the clustering cluster], Output: Only the text of the general label". Label verification means that if the label output by the large model exceeds 6 characters or cannot cover all summaries in the cluster, correct it through a secondary prompt (such as "Please streamline the label to within 6 characters"). The output refers to assigning a unique standardized general label to each valid cluster, forming a "cluster-label" mapping relationship. The format can be {cluster ID: standardized general label, list of comment IDs within the cluster, list of summary points of issues within the cluster}. This step automatically generates labels using a large model, solving the problems of "low efficiency and strong subjectivity in manual label naming" in traditional methods. It also ensures a high degree of label standardization (only one label for the same type of problem), avoiding label fragmentation.

[0024] S7. Tag Back-injection and Association: Establish the mapping relationship between the standardized general tags generated in step S6 and the valid original user comments in step S1, forming a tagged user comment dataset; Specifically, a strong association is established between the "standardized general tags" generated in step S6 and the "original user comments" in step S1. That is, the tags are back-mapped to the original comments through the "comment ID" to form a "tagged user comment dataset" - this dataset contains both the original comment text and the corresponding issue tags, supporting subsequent "issue tracing" (finding the original comments by tags) and "quantitative statistics" (statistical analysis of frequency by tags). The specific operations include: Association mapping: Based on the "cluster-comment ID" relationship in step S5 and the "cluster-tag" relationship in step S6, the "comment ID-tag" mapping is obtained by concatenating the comment IDs; Dataset Construction: Integrate raw comment information and tags to form a structured dataset, which can be formatted as {Comment ID: [Standardized general tag, original comment text, sentiment dimension, publication time, data source]}; Data storage: Store the dataset in CSV or JSON format for easy access by subsequent statistical analysis tools (such as Python Pandas, Tableau); This step achieves "binding tags to original comments," solving the shortcomings of traditional methods where "tags and comments are disconnected and cannot be traced back to their source." Businesses can quickly retrieve all relevant original comments through tags, gain a deeper understanding of the specific details of the problem, and determine the specific issue.

[0025] S8. Tag statistical analysis: Based on the tagged user comment dataset from step S7, perform tag frequency statistics, time trend analysis, and problem source tracing query. In step S8, the problem source tracing query is performed by inputting a standardized general tag and retrieving all valid original user comment texts corresponding to that standardized general tag. Specifically, based on the "labeled user review dataset" from step S7, multi-dimensional quantitative analysis is conducted to provide data support for enterprise decision-making. The core analysis includes: a. Tag frequency statistics Calculate the frequency and percentage of each "standardized general tag" in all negative comments to identify "high-frequency core issues". The calculation formula is: Tag percentage = (Number of comments corresponding to this tag / Total number of negative comments) × 100%.

[0026] b. Time trend analysis Analyze the frequency changes of each tag by time unit (day, week, month) to identify "problem trends" (e.g., rising, falling, or stable). For example, the daily frequency of "charging overheating problem" is as follows: 15 cases on January 10, 20 cases on January 11, 28 cases on January 12, and 35 cases on January 13, showing an overall upward trend, indicating that the problem may be escalating and requires urgent investigation (e.g., whether it is a quality issue with a particular batch of products). The frequency of "headphone disconnection issues" was analyzed weekly: 120 cases in week 1, 80 cases in week 2, and 50 cases in week 3, showing an overall downward trend, indicating that the issue has been resolved through product upgrades.

[0027] c. Problem tracing and inquiry It provides a quick query function for "tags-original comments": by entering any standardized general tag, you can retrieve all the original comment text, posting time, and data source corresponding to that tag, helping companies to analyze the details of the problem in depth; d. Dimensional cross-analysis By combining fields such as "emotional dimension" and "data source", cross-analysis can be performed to pinpoint the specific scenario of the problem.

[0028] This step transforms "problem labels" into decision-making "data insights" through multi-dimensional statistics and source tracing, thus addressing the shortcomings of traditional methods where "mining results cannot support quantitative decision-making."

[0029] This embodiment does not impose any limitation on the shape, material, structure, etc. of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for mining open-ended questions based on a large model for user reviews, characterized in that: Includes the following steps: S1. User comment input and preprocessing: Obtain the original user comment data source and preprocess the original user comment data source to obtain valid original user comments; S2. Fine-grained sentiment analysis of large model: Call the pre-trained large model to perform sentiment analysis on each of the valid original user comments obtained in step S1, and output the sentiment polarity and sentiment dimension of each valid original user comment. The sentiment polarity includes negative and non-negative. S3. Open-ended question summary generation: Select valid original user comments with negative sentiment polarity from step S2 as target comments, and call the pre-trained large model to generate question summary for each target comment through prompting engineering; S4. Question point sentence vector encoding: Use a pre-trained sentence vector model to convert all the question point summaries generated in step S3 into semantic vectors; S5. Problem point clustering and noise reduction: The DBSCAN clustering algorithm is used to cluster the semantic vectors obtained in step S4, merge semantically similar problem point summaries to form clusters, and automatically remove noise points; S6. Automated generation of cluster labels: The pre-trained large model is called to process the problem point summaries in each cluster obtained in step S5 and generate standardized general labels. S7. Tag Back-injection and Association: Establish the mapping relationship between the standardized general tags generated in step S6 and the valid original user comments in step S1, forming a tagged user comment dataset; S8. Tag Statistical Analysis: Based on the tagged user review dataset from step S7, perform tag frequency statistics, time trend analysis, and problem source tracing query.

2. The open-ended question point mining method based on a large model for user reviews according to claim 1, characterized in that: The original user review data source in step S1 includes at least one of e-commerce platform reviews, social media texts, and customer service conversation records. The preprocessing includes removing duplicate reviews, filtering meaningless text, and assigning a unique review ID to each valid original user review.

3. The open-ended question point mining method based on a large model for user reviews according to claim 1, characterized in that: In step S2, non-negative includes positive and neutral, and the emotional dimension includes at least one of product quality, service attitude and user experience. In step S2, the pre-trained large model is GPT-4 or Claude 3.

5.

4. The open-ended question point mining method based on a large model for user reviews according to claim 1, characterized in that: In step S3, the problem summary is a core complaint description of no more than 10 words, without emotional words or irrelevant details.

5. The open-ended question point mining method based on a large model for user reviews according to claim 1, characterized in that: In step S8, the problem tracing query involves retrieving all valid original user comments corresponding to a standardized general tag by inputting that tag.

6. The open-ended question point mining method based on a large model for user reviews according to claim 1, characterized in that: The pre-trained sentence vector model in step S4 can be either Sentence-BERT or Connan-embedding.

Images