Refined user viewpoint extraction and analysis method based on large model

Abstract

The invention relates to the technical field of natural language processing, and discloses a large model-based refined user viewpoint extraction and analysis method, which comprises the following steps of: obtaining user comment data; analyzing the user comment data by using a large model, and identifying viewpoint fragments; on the basis of the recognized viewpoint fragments, the following processing is executed by using a large model, and the processing steps do not have a fixed sequence and are allowed to be performed in parallel or asynchronously; based on the identified at least one problem point, utilizing a large model to automatically generate at least one new label, and adding the at least one new label to an existing label system so as to realize expansion and updating of the label system; and generating a six-tuple analysis result of the at least one tag. According to the method, question points in user comments are recognized in real time through a large model, new changes and emerging demands fed back by users are captured and responded in time, it is ensured that the analysis result is always kept synchronous with the hearts and sounds of the users, and insight with higher timeliness and guiding significance is provided for enterprises.

Description

Technical Field

[0001] The present invention relates to a large-model-based refined user opinion extraction and analysis method, belonging to the technical field of text mining and user opinion analysis in natural language processing. Background Art

[0002] With the rapid development and popularization of Internet technology, the comment data generated by users through various online platforms (such as e-commerce websites, social media, app stores, etc.) has exploded. These massive user comments contain valuable information such as product feedback, user needs, and market trends, which are crucial for companies to improve their products and services and formulate effective market strategies. Therefore, how to efficiently and accurately extract valuable user opinions from these unstructured text data and conduct in-depth analysis has become a technical issue of widespread concern in the current industry.

[0003] At present, the most common method in the industry for analyzing user comments is based on the traditional tag system. This method usually requires manual definition of a series of product or service-related tags in advance, and then classifies user comments into preset tags through technical means such as text classification and keyword matching, so as to achieve statistics and analysis of user feedback. For example, for a mobile phone product, labels such as battery life, camera effect, screen display, etc. may be pre-defined, and then user comments are classified according to the content mentioned. However, the user comment analysis method based on the traditional tag system has exposed many limitations in practical applications: 1. Once the traditional label system is established, its update and expansion often requires a lot of manpower and time. When the market environment changes, product functions are updated or new user concerns emerge, the original label system is difficult to cover these new information, resulting in the analysis results being unable to timely reflect the latest user feedback and potential needs. For example, when a new mobile phone function is released, users may generate new comments about the function, and the pre-set labels may not be able to accurately capture these new views.

[0004] 2. Traditional labeling systems can usually only provide relatively rough classification and statistical results, and it is difficult to dig deeper into more detailed and specific opinions in user comments. For example, a user may comment that the photo effect is good, but traditional labels may only classify it under the photo effect label, and cannot further distinguish whether the photo effect is good during the day or at night, nor can it identify more specific evaluation dimensions such as user focus speed and color reproduction. In addition, traditional methods often find it difficult to accurately distinguish the subject and object of the evaluation. For example, if a user comments that the battery of this phone is much better than my previous phone, traditional methods may find it difficult to clearly identify that the evaluation subject is the battery of this phone and the comparison object is the previous phone.

[0005] 3. The establishment and maintenance of the label system requires manual definition and adjustment by domain experts, which is costly and inefficient. With the continuous growth of user review data and the diversity of user language expressions, it is difficult for manually maintained label systems to keep up with the pace of data changes and to achieve large-scale automated analysis and expansion.

[0006] In order to overcome the shortcomings of the traditional labeling system, some user comment analysis methods based on machine learning and natural language processing technologies have emerged in recent years. These methods attempt to automatically identify key information in user comments through algorithms, such as sentiment analysis and keyword extraction. However, these methods still have certain limitations when dealing with complex user views and segmented scenarios, such as difficulty in accurately understanding the deep meaning of user comments, inability to provide structured analysis results, and difficulty in realizing the automated construction and evolution of the labeling system.

[0007] Therefore, how to construct a user opinion extraction and analysis method that can dynamically adapt to changes in user feedback, provide multi-dimensional refined analysis, and have high efficiency and low manual dependence has become a technical problem to be solved by the present invention. Summary of the invention

[0008] The present invention provides a method for extracting and analyzing refined user opinions based on a large model, the main purpose of which is to solve the problems of insufficient flexibility, difficulty in conducting in-depth and fine-grained analysis, and high dependence on manual work in traditional labeling systems in user comment analysis.

[0009] To achieve the above object, the present invention provides a method for extracting and analyzing user opinions in a refined manner based on a large model, comprising the following steps: Step 1, obtain user comment data; Step 2, using a large model to analyze the user comment data and identify opinion fragments; Step 3, based on the identified opinion fragments, use the big model to perform the following processing, and there is no fixed order between the processing steps, and parallel or asynchronous processing is allowed: Sub-step 3.1, summarize at least one problem point reflected in the opinion fragment; Sub-step 3.2, extract the evaluation subject and evaluation object in the opinion fragment; Sub-step 3.3, subdivide the usage scenarios involved in the opinion fragment to obtain at least one subdivided scenario; Step 4: based on the at least one problem point identified, automatically generate at least one new label using the large model, and add the at least one new label to the existing label system to achieve expansion and update of the label system; Step 5, generating a six-tuple analysis result including the evaluation subject, the evaluation object, the opinion fragment, the at least one segmented scenario, the at least one problem point, and at least one tag related to the user comment data in the expanded and updated tag system.

[0010] Preferably, the large model is a pre-trained language model based on the Transformer architecture.

[0011] Preferably, the step of identifying and summarizing at least one problem point reflected in the user comments includes: performing semantic analysis on the user comment data to extract the core demands expressed by the user and the specific difficulties encountered; summarizing and abstracting the extracted core demands and specific difficulties to form a concise and clear problem description.

[0012] Preferably, the step of automatically generating at least one new label through the large model based on the at least one identified problem point includes: analyzing the semantic features and contextual information of the at least one problem point; and generating a new label that can accurately summarize the at least one problem point based on the semantic features and contextual information.

[0013] Preferably, the step of using the large model to analyze the user comment data, identify opinion fragments, and extract evaluation subjects and evaluation objects includes: performing dependency syntactic analysis or semantic role labeling on the user comment data; determining, based on the analysis results, sentence fragments expressing user evaluations as opinion fragments; identifying in the opinion fragments the subject that issues the evaluation action as the evaluation subject, and the object to which the evaluation action is directed as the evaluation object.

[0014] Preferably, the step of using the big model to segment the usage scenarios mentioned in the user review data to obtain at least one segmented scenario includes: identifying descriptions in user reviews related to the way, environment or purpose of product or service usage; clustering or classifying the related descriptions to obtain segmented scenarios with clear semantic orientation.

[0015] Preferably, before the step of acquiring user comment data, the method further includes a step of preprocessing the user comment data, wherein the preprocessing includes removing at least one of noise data, irrelevant information, punctuation marks and stop words.

[0016] Preferably, the method is applied to at least one of the following fields: product evaluation analysis, competitive product analysis, social media monitoring and analysis, customer service and feedback processing, or market research and trend forecasting in the e-commerce industry.

[0017] Preferably, the method further comprises the step of using the tags in the expanded and updated tag system to further analyze and count the user comment data to generate a quantitative analysis report.

[0018] Preferably, the step of automatically generating at least one new label and adding the at least one new label to the existing label system is performed periodically, or is triggered to be performed when a preset number of new problem points are identified.

[0019] Compared with the problems described in the background technology, the beneficial effects of the present invention are: 1. The user comment problem points identified by the big model are directly used to drive the dynamic evolution of the label system, breaking away from the traditional label system's reliance on predefined rules and building an insight system that can self-optimize as user feedback content changes. This inherent adaptability enables companies to more keenly capture users' real pain points and emerging needs, thereby avoiding missing key information due to solidified labels.

[0020] 2. By using a large model to divide user comment scenarios in detail and combining them with dynamically updated tags, this method can achieve scenario-based in-depth analysis of user feedback. This fusion of multi-dimensional information can not only identify what the user said, but also understand the context in which the user said it, thereby helping companies to more accurately understand the root causes of the problem and provide more targeted clues for product optimization and service innovation.

[0021] 3. Traditional methods often passively rely on preset labels to analyze user feedback, while this method gives enterprises greater initiative. Through problem-driven label system updates, enterprises can dynamically adjust the analysis framework based on actual user feedback, more flexibly focus on the most critical issues at the moment, and thus respond more efficiently to market changes and user needs.

[0022] 4. The six-tuple analysis results generated by this method are not only a structured presentation of user feedback, but also a knowledge base that is continuously learning and evolving. With the dynamic update of the tag system, subsequent user review analysis will be able to be conducted based on richer and more accurate tags, continuously exploring and releasing the potential value of user feedback, and providing continuous decision support for enterprises.

[0023] 5. In the context of rapid product iteration and ever-changing market environment, the dynamic adaptability and intelligent analysis capabilities of this method can help companies quickly understand user feedback on new features and new products, identify potential problems in a timely manner and make adjustments, thereby maintaining a leading position in the fierce market competition and better meeting users' ever-changing needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1This is a flow chart of the refined user opinion extraction and analysis method based on a large model of the present invention.

[0025] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0026] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0027] The present application embodiment provides a method for extracting and analyzing user opinions in a refined manner based on a large model, comprising the following steps: Step 1, obtain user comment data; Step 2, using a large model to analyze the user comment data and identify opinion fragments; Step 3, based on the identified opinion fragments, use the big model to perform the following processing, and there is no fixed order between the processing steps, and parallel or asynchronous processing is allowed: Sub-step 3.1, summarize at least one problem point reflected in the opinion fragment; Sub-step 3.2, extract the evaluation subject and evaluation object in the opinion fragment; Sub-step 3.3, subdivide the usage scenarios involved in the opinion fragment to obtain at least one subdivided scenario; Step 4: based on the at least one problem point identified, automatically generate at least one new label using the large model, and add the at least one new label to the existing label system to achieve expansion and update of the label system; Step 5, generating a six-tuple analysis result including the evaluation subject, the evaluation object, the opinion fragment, the at least one segmented scenario, the at least one problem point, and at least one tag related to the user comment data in the expanded and updated tag system.

[0028] Preferably, the large model is a pre-trained language model based on the Transformer architecture.

[0029] Preferably, the step of identifying and summarizing at least one problem point reflected in the user comments includes: performing semantic analysis on the user comment data to extract the core demands expressed by the user and the specific difficulties encountered; summarizing and abstracting the extracted core demands and specific difficulties to form a concise and clear problem description.

[0030] Preferably, the step of automatically generating at least one new label through the large model based on the at least one identified problem point includes: analyzing the semantic features and contextual information of the at least one problem point; and generating a new label that can accurately summarize the at least one problem point based on the semantic features and contextual information.

[0031] Preferably, the step of using the large model to analyze the user comment data, identify opinion fragments, and extract evaluation subjects and evaluation objects includes: performing dependency syntactic analysis or semantic role labeling on the user comment data; determining, based on the analysis results, sentence fragments expressing user evaluations as opinion fragments; identifying in the opinion fragments the subject that issues the evaluation action as the evaluation subject, and the object to which the evaluation action is directed as the evaluation object.

[0032] Preferably, the step of using the big model to segment the usage scenarios mentioned in the user review data to obtain at least one segmented scenario includes: identifying descriptions in user reviews related to the way, environment or purpose of product or service usage; clustering or classifying the related descriptions to obtain segmented scenarios with clear semantic orientation.

[0033] Preferably, before the step of acquiring user comment data, the method further includes a step of preprocessing the user comment data, wherein the preprocessing includes removing at least one of noise data, irrelevant information, punctuation marks and stop words.

[0034] Preferably, the method is applied to at least one of the following fields: product evaluation analysis, competitive product analysis, social media monitoring and analysis, customer service and feedback processing, or market research and trend forecasting in the e-commerce industry.

[0035] Preferably, the method further comprises the step of using the tags in the expanded and updated tag system to further analyze and count the user comment data to generate a quantitative analysis report.

[0036] Preferably, the step of automatically generating at least one new label and adding the at least one new label to the existing label system is performed periodically, or is triggered to be performed when a preset number of new problem points are identified.

[0037] Example 1: Assume that an e-commerce platform that sells smart home products has received a large number of user reviews after launching a new smart sweeping robot. These reviews include user feedback on various aspects of the product, such as product functions, performance, and user experience. The platform hopes to quickly and accurately understand users' evaluation of this new product, especially to discover potential problems and new user needs, so as to make timely product improvements and optimizations.

[0038] For example, after users purchased the smart sweeping robot on an e-commerce platform and used it for a period of time, they published their experience and evaluation through the comment system provided by the platform. These comments are presented in the form of natural language text, covering the product's functions (such as cleaning ability, obstacle avoidance ability, App control, etc.), performance (such as battery life, noise level, etc.), user experience (such as ease of operation, degree of intelligence, etc.), as well as the problems encountered by users and the suggestions they put forward. The platform first collects all the user's comment data on the smart sweeping robot. Then, the data is preprocessed, including removing HTML tags, special symbols, irrelevant advertising information, etc. in the comments, and performing word segmentation processing to prepare for the subsequent large model analysis. For example, an original comment may be: This sweeping robot cleans very cleanly, but the App is a bit difficult to use, and I hope it can be optimized later. After preprocessing, you may get something like [clean][clean][but][App][a bit][difficult to use][hope][later][can][optimize].

[0039] The preprocessed comment data is input into the pre-trained big model. The big model identifies the opinions and potential problem points expressed in user comments through semantic understanding of the text. For example, for a comment that the App is a bit difficult to use and hopes to be optimized later, the big model may identify the problem point as the App is not convenient to operate. The system then compares this newly identified problem point, "The App is not convenient to operate," with the existing labeling system. If there is no similar label in the existing labeling system (for example, there may only be broader labels such as "App function is missing"), the big model will automatically generate a new, more precise label, such as "App is not convenient to operate," and add it to the labeling system.

[0040] The large model continues to analyze user comments, identifies fragments that express opinions, and extracts the subject and object of the evaluation. For the comment that this sweeping robot cleans very well, but the App is a bit difficult to use, the large model will identify two opinion fragments: it cleans very well and the App is a bit difficult to use. At the same time, the evaluation subjects are extracted as this sweeping robot and the App. The large model also analyzes the usage scenarios mentioned in the comments. For example, the user commented that the cleaning effect on wooden floors is very good, but the suction is not enough on carpets. The large model will identify two sub-scenarios: wooden floors and carpets.

[0041] Finally, for each user comment, the system will generate a structured analysis result containing the following six elements, including: subject: for example, this sweeping robot, App; object: in this example, it may be empty, or if the comment is comparative, the comparative object will be extracted; opinion fragment: for example, it cleans very well, it is a bit difficult to use, and the suction is not enough; subdivided scenes: for example, wooden floors, carpets; problem points: for example, the App is not convenient to operate, and the suction on the carpet is insufficient; tags: for example, the cleaning effect is good, the App is not convenient to operate, the carpet has insufficient suction, and the battery life is satisfactory (assuming that other comments have generated this tag), etc. (Among them, the App is not convenient to operate is a tag newly added in this dynamic update). If a large number of users report that the new version of the App is inconvenient to operate, through problem point identification and dynamic tag generation, the platform can quickly discover that the App is not convenient to operate and has become a high-frequency problem, without waiting for manual definition and statistics of this new tag. This enables the platform to respond to user feedback more quickly and optimize the App in a timely manner.

[0042] For example, through the association between the carpet segment and the insufficient suction opinion fragment, the platform can clearly understand the user's dissatisfaction with the product performance in a specific scenario, so as to make more targeted product improvements, such as enhancing the suction of the sweeping robot on the carpet. The structured analysis results of the six-tuple not only include the user's emotional tendencies, but also include the object of the evaluation, the scene where it occurred, and the specific problem points, providing the platform with richer and more detailed user feedback information, helping it to understand the advantages and disadvantages of the product from multiple dimensions. For example, the platform may find that users are generally satisfied with the cleaning ability and battery life of the sweeping robot (through high-frequency positive reviews and related tags), but there are many complaints about the ease of use of the App (through the newly generated App operation is not convenient label and related negative opinion fragments). The traditional labeling system requires manual monitoring of user comments, discovering new problems and manually adding new labels. In this embodiment, the large model can automatically complete these tasks, greatly reducing labor costs, and improving analysis efficiency and label system coverage. Through refined analysis and label management of user comments, the platform can better understand user preferences and needs, and provide more accurate data support for subsequent personalized recommendations, precision marketing, and user portrait construction. For example, for users who often mention pet hair cleaning, the platform can more specifically recommend a robot vacuum cleaner model with a powerful vacuuming function.

[0043] Example 2: In order to better understand users' feedback on the new short video editing function on its platform and to promptly discover and respond to possible negative public opinions, a social media platform adopted the large model-based refined user opinion extraction and analysis method of the present invention.

[0044] For example, on this social media platform, users can post their experience, opinions and suggestions on various features of the platform. When the platform releases a new short video editing function, users will generate a large number of text comments when posting short videos or under related topics. The platform needs to monitor these comments in real time to understand users' acceptance of the new function, existing problems and potential public opinion risks.

[0045] The platform captures the comments data of users on social media that contain keywords related to the new short video editing function in real time. The data is cleaned and pre-processed to remove irrelevant information, advertising content, and user privacy information to ensure data quality. For example, a user comment might be: This new editing function is great, the special effects are cool, but it is a bit slow to export the video.

[0046] The preprocessed comment data is input into the big model, which identifies snippets of opinions expressed by users, such as the editing function is great, the special effects are cool, and the video export is a bit slow. At the same time, the big model summarizes potential problem points, such as slow video export. The system compares the identified problem point, slow video export, with the existing label system. If there are no sufficiently detailed labels in the existing label system (for example, there may only be problems with the video function, etc.), the big model will automatically generate more specific labels, such as slow video export, and add them to the label system. The big model recognizes that the evaluation subject is this new editing function and exported video. Due to the characteristics of social media, the usage scenarios may be relatively broad, such as daily sharing, Vlog production, etc. The big model will infer and divide according to the content of the comments.

[0047] Generate six-tuple analysis results: For each comment, generate a six-tuple analysis result containing the following elements, including subject: for example, this new editing function, export video; object: none or the platform itself; opinion fragment: for example, it’s great, the special effects are cool, and it’s a bit slow; segmented scenarios: for example, daily sharing, Vlog production; problem points: for example, the video export speed is slow; tags: for example, the editing function is well received, the special effects are cool, the video export speed is slow, the user experience is smooth (assuming that other comments have generated this tag), etc.

[0048] Through this implementation method, the platform can monitor users' evaluation of short video editing functions in real time, especially potential negative problems such as slow video export speed can be quickly identified and marked, which helps the platform take timely measures for optimization. Through dynamically generated tags and problem point summaries, the platform can quickly discover the problems that users have concentrated on, such as slow video export speed may become a high-frequency tag, helping the platform to timely warn and respond to possible negative public opinion to avoid the spread of problems. By analyzing the combination of opinion fragments, segmented scenes and tags, the platform can understand which special effects users prefer and in which scenarios they use new functions more often, thereby providing data support for subsequent function iterations and user operations. Compared with manual monitoring and monitoring methods based on fixed keywords, this method uses the semantic understanding ability of large models to more accurately identify users' true intentions and potential negative emotions, and automatically updates the tag system, improving the efficiency and accuracy of public opinion monitoring; through refined analysis of user feedback, the platform can clearly understand users' specific needs and dissatisfaction with new functions, such as users want to increase the speed of video export, which provides a clear optimization direction for the product development team.

[0049] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A refined user opinion extraction and analysis method based on a large model, characterized in that: The following steps are involved: Step 1, obtain user comment data; Step 2, using a large model to analyze the user comment data and identify opinion fragments; Step 3, based on the identified opinion fragments, use the big model to perform the following processing, and there is no fixed order between the processing steps, and parallel or asynchronous processing is allowed: Sub-step 3.1, summarize at least one problem point reflected in the opinion fragment; Sub-step 3.2, extract the evaluation subject and evaluation object in the opinion fragment; Sub-step 3.3, subdivide the usage scenarios involved in the opinion fragment to obtain at least one subdivided scenario; Step 4: based on the at least one problem point identified, automatically generate at least one new label using the large model, and add the at least one new label to the existing label system to achieve expansion and update of the label system; Step 5, generating a six-tuple analysis result including the evaluation subject, the evaluation object, the opinion fragment, the at least one segmented scenario, the at least one problem point, and at least one tag related to the user comment data in the expanded and updated tag system.

2. The method for extracting and analyzing user opinions based on a large model according to claim 1 is characterized in that: The large model is a pre-trained language model based on the Transformer architecture.

3. The method for extracting and analyzing user opinions in a refined manner based on a large model according to claim 1 is characterized in that: The step of identifying at least one problem point includes: performing semantic analysis on the user comment data to extract the core demands expressed by the user and the specific difficulties encountered; summarizing and abstracting the extracted core demands and specific difficulties to form a concise and clear problem description.

4. The method for extracting and analyzing user opinions based on a large model according to claim 1 is characterized in that: Based on the identified at least one problem point, the step of automatically generating at least one new label through the large model includes: analyzing the semantic features and context information of the at least one problem point; generating a new label that can accurately summarize the at least one problem point based on the semantic features and context information.

5. The method for extracting and analyzing user opinions in a refined manner based on a large model according to claim 1 is characterized in that: The steps of using the large model to analyze the user comment data, identify opinion fragments, and extract evaluation subjects and evaluation objects include: performing dependency syntactic analysis or semantic role labeling on the user comment data; determining sentence fragments expressing user evaluations as opinion fragments based on the analysis results; and identifying the subject that issues the evaluation action in the opinion fragment as the evaluation subject, and the object to which the evaluation action points as the evaluation object.

6. The method for extracting and analyzing user opinions in a refined manner based on a large model according to claim 1 is characterized in that: The step of using the big model to segment the usage scenarios mentioned in the user review data to obtain at least one segmented scenario includes: identifying descriptions in user reviews related to the way, environment or purpose of using the product or service; clustering or classifying the related descriptions to obtain segmented scenarios with clear semantic orientation.

7. The method for extracting and analyzing user opinions in a refined manner based on a large model according to claim 1 is characterized in that: Before the step of acquiring user comment data, the method further includes a step of preprocessing the user comment data, wherein the preprocessing includes removing at least one of noise data, irrelevant information, punctuation marks and stop words.

8. The method for extracting and analyzing user opinions in a refined manner based on a large model according to claim 1 is characterized in that: The method is applied to at least one of the following fields: commodity evaluation analysis, competitive product analysis, social media monitoring and analysis, customer service and feedback processing, or market research and trend forecasting in the e-commerce industry.

9. The method for extracting and analyzing user opinions in a refined manner based on a large model according to claim 1 is characterized in that: The method also includes the step of further analyzing and counting the user comment data using the tags in the expanded and updated tag system to generate a quantitative analysis report.

10. The method for extracting and analyzing user opinions in a refined manner based on a large model according to claim 1, characterized in that: The step of automatically generating at least one new label and adding the at least one new label to the existing label system is performed periodically, or is triggered to be performed when a preset number of new problem points are identified.

Images