APP software defect identification method based on user comments

Through the adaptive multi-dimensional feature dynamic network, user comment data is processed and defects of APP software are identified, which solves the shortcomings of traditional testing methods in covering user problems and efficiency, and achieves more efficient and accurate software defect identification and quality assurance.

CN118966167AInactive Publication Date: 2024-11-15CHONGQING UNIV OF POSTS & TELECOMM

Patent Information

Application Number
CN202410943682.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Traditional software testing methods are difficult to fully cover the software problems encountered by users in actual use, and manual analysis of user feedback is time-consuming and inefficient, so software defects cannot be effectively identified.

Method used

Adaptive multidimensional feature dynamic network method based on user comments is adopted, and the context range is dynamic and context is dynamically adjusted by obtaining and preprocessing user comment data, important keyword scores are calculated, multi-dimensional features are constructed, and multi-dimensional feature dynamic network is used to identify defects, and context range is dynamically adjusted. Network structure.

Benefits of technology

Improves the efficiency and accuracy of software defect identification, captures subtle defects that traditional testing may miss, provides a more comprehensive quality assurance tool, and maintains high performance in dynamic environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118966167A_ABST
    Figure CN118966167A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of natural language processing, and particularly relates to an APP software defect identification method based on user comments. The method comprises the following steps: acquiring comment data of a user on software and preprocessing the comment data to obtain preprocessed user comment data; calculating important scores of all keywords in the user comment data; selecting n keywords with the highest importance score, and constructing multi-dimensional features of comments according to the n keywords and the user comment data; the trained self-adaptive multi-dimensional feature dynamic network is adopted to process the multi-dimensional features, and an APP software defect recognition result is obtained; according to the method, the efficiency of identifying software problems can be improved, and tiny defects which may be omitted in a traditional test can be captured, so that a more comprehensive quality assurance tool is provided for software developers.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and specifically relates to an APP software defect identification method based on user comments. Background Art

[0002] After more than a decade of development, mobile applications (Apps) have become increasingly indispensable in people's daily lives. Developers have developed millions of available apps to handle a variety of tasks (such as shopping, entertainment, and social interaction). The importance of apps has prompted development teams to understand the functional requirements and bug reports from users, and to carry out quality assurance and software maintenance activities. App users usually write reviews for the apps they use on platforms such as app stores, certain specialized technical forums, or personal blogs. These user reviews are usually short texts that can provide App developers with valuable information, such as user experience, bug reports, and requirements for new features. Fully understanding these reviews can help developers improve App quality and user satisfaction with their products.

[0003] In traditional software testing, testers or automated test scripts often look for defects during the development process, but these methods may not fully cover all the problems users encounter in actual use. In addition, manually analyzing a large amount of user feedback is time-consuming and inefficient. Therefore, the development of automated tools and methods to be able to extract useful information from large-scale user data has become the key to improving software quality and user satisfaction.

[0004] In summary, there is an urgent need for a method that can automatically extract useful information from user comments to identify software defects, which not only improves the efficiency of identifying software problems, but also captures subtle defects that may be missed by traditional testing, thereby providing software developers with a more comprehensive quality assurance tool. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes an APP software defect identification method based on user comments, which includes:

[0006] S1: Obtain user review data on the software and preprocess it to obtain preprocessed user review data;

[0007] S2: Calculate the importance scores of all keywords in the user review data;

[0008] S3: Select n keywords with the highest importance scores, and construct multi-dimensional features of the comments based on the n keywords and user comment data;

[0009] S4: Use the trained adaptive multi-dimensional feature dynamic network to process the multi-dimensional features and obtain the APP software defect recognition results.

[0010] Preferably, the process of preprocessing the comment data includes:

[0011] The comment data is cleaned to remove irrelevant information to obtain the cleaned comment data; the cleaned comment data is cut into words with clear meanings to obtain the word segmentation result, that is, the pre-processed user comment data.

[0012] Preferably, the process of calculating the importance score of a keyword includes:

[0013] Perform word frequency statistics on the words in the comment data, set a word frequency threshold, and use words that exceed the threshold as keywords;

[0014] Set the co-occurrence window size, record the number of co-occurrences of each keyword with other keywords in the co-occurrence window; calculate the degree centrality of keywords based on the number of co-occurrences of keywords;

[0015] Assign initial weights to keywords based on their degree centrality;

[0016] Calculate the average position of each keyword in the comments and the average position of all keywords; adjust the initial weight of the keyword according to the average position of the keyword in the comments and the average position of all keywords to obtain the final weight of the keyword;

[0017] The final weights of the keywords are normalized to obtain the importance scores of the keywords.

[0018] Furthermore, the formula for adjusting the initial weight of keywords is:

[0019]

[0020] Among them, W final (k) represents the final weight of keyword k, W initial (k) represents the initial weight of keyword k, p k represents the average position of keyword k, μ represents the average position of all keywords, and δ represents the standard deviation of keyword positions.

[0021] Preferably, the process of constructing multi-dimensional features includes:

[0022] S31: Calculate the semantic similarity between each keyword and other keywords respectively; take the average of all semantic similarities to obtain the semantic similarity score of the keyword;

[0023] S32: Obtain the context of all keywords, and calculate the context similarity between the context of each keyword and the context of other keywords respectively; take the average of all context similarities to obtain the context similarity score of the keyword;

[0024] S33: performing weighted summation of the semantic similarity score and the context similarity score of the keyword to obtain a comprehensive score of the keyword;

[0025] S34: expanding or reducing the context of the keyword according to the comprehensive score of the keyword, and using the context of the keyword as the text feature of the comment in which it is located;

[0026] S35: Obtain the publishing time of the comment containing the keyword and use it as a time feature;

[0027] S36: Calculate the user score of the comment according to the community feedback of the comment where the keyword is located, input the comment where the keyword is located into the pre-trained LSTM model to obtain the sentiment score of the comment; obtain the length of the comment where the keyword is located;

[0028] S37: Combine the text features, time features, comment length, user scores and sentiment scores of the comments to obtain multi-dimensional features.

[0029] Furthermore, the process of expanding or reducing the context of the keyword includes: setting a comprehensive score threshold; if the comprehensive score of the keyword is greater than the comprehensive score threshold, expanding the context range of the keyword by 2 words before and after, otherwise, reducing the context range of the keyword by 2 words before and after; repeatedly calculating the comprehensive score of the keyword, and expanding or reducing the context of the keyword until the context range reaches a maximum value or the comprehensive score of the keyword is stable.

[0030] Furthermore, the process of calculating the user score of the comment includes:

[0031] Assign weights to the three indicators, namely the number of likes, the number of citations, and the number of replies; perform weighted summation based on the number of likes, the number of citations, and the number of replies of the comment and the weights of their corresponding indicators to obtain the initial user score; normalize the initial user score to obtain the final user score.

[0032] Preferably, during the training process of the adaptive multi-dimensional feature dynamic network, the MDEM evaluation mechanism is used to verify the network training effect and adjust the network parameters according to the feedback results of the MDEM evaluation mechanism; during training, the adaptive multi-dimensional feature dynamic network optimizes the network structure according to the training effect, and dynamically adjusts the learning rate according to the training error, verification error and the speed of parameter change.

[0033] Furthermore, the formula of the MDEM evaluation mechanism is:

[0034]

[0035] Among them, S represents the comprehensive score of the model, N represents the number of test sets, and Accuracy i represents the accuracy on the i-th test set; K represents the number of cross-validation folds, F1-Score k,cross-val represents the F1 score of the k-fold cross validation; Var(Predictions) and Var(Inputs) represent the variance of the prediction results and the variance of the input data respectively; M represents the number of different conditions; Error j,stability represents the model stability difference under the jth condition; J is the number of consistency tests, Consistency j represents the consistency score of the jth test; L is the number of model complexity factors considered, Complexity l represents the measure of the lth complexity factor; α, β, γ, δ,∈, ζ represent the first to sixth weight factors respectively.

[0036] Furthermore, the formula for optimizing the network structure according to the training effect is:

[0037]

[0038] Among them, Θ(l t ,E) represents the optimized network structure, l t represents the current hierarchical structure, E represents the overall training effect of the adaptive multi-dimensional feature dynamic network, T represents the time window size considered, and e t represents the training error at time point t, v t represents the validation error at time point t, l t-i represents the current i-th hierarchy, N represents the number of historical layers considered; φ, ω, ξ, η represent the first to fourth adjustment coefficients respectively.

[0039] The beneficial effects of the present invention are:

[0040] 1. Most traditional methods have a fixed range when processing keyword contexts and cannot be adjusted dynamically, which may result in insufficient or excessive context information, affecting recognition accuracy. The present invention introduces a comprehensive score calculation mechanism to dynamically adjust the context range based on the importance of keywords. This adaptive mechanism enables keywords to obtain the most appropriate context information in different situations, improving the accuracy of defect recognition.

[0041] 2. Most traditional methods rely on static feature extraction and fixed-structure neural networks, and cannot dynamically adjust the network structure to adapt to data changes. Adaptive multi-dimensional feature dynamic networks can dynamically adjust the number of network layers and neurons according to the complexity of the data and the training effect, ensuring that the network can maintain optimal performance when processing data of different scales and complexities. This flexible network structure greatly improves the generalization ability and recognition accuracy of the model.

[0042] 3. The present invention constructs a multi-dimensional feature system including text features, time features, comment length, user scores and sentiment scores, which fully captures various information in the comment data, thereby more comprehensively identifying software defects. The comprehensive use of this multi-dimensional feature makes the recognition results more comprehensive and accurate.

[0043] 4. By comprehensively considering multiple factors such as semantic similarity and contextual similarity, the present invention can effectively capture subtle defects hidden in user comments, which are often easily missed by traditional testing methods. This capability enables the present invention to provide a more comprehensive and sophisticated quality assurance tool.

[0044] 5. The present invention utilizes the AMFDN model to continuously update and adjust as new data is added, maintaining efficient problem identification capabilities. This continuous self-improvement capability enables the present invention to maintain high performance in a dynamically changing environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of the APP software defect identification method based on user comments in the present invention;

[0046] Figure 2 This is a training diagram of the AMFDN network in the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] The present invention proposes an APP software defect identification method based on user comments, such as Figure 1 As shown, the method includes:

[0049] S1: Obtain user review data on the software and preprocess it to obtain preprocessed user review data.

[0050] User review data of software can be obtained from channels such as APP application stores. The user review data includes text content (review subject, words), review time, review and sentiment (user rating, sentiment tendency), problem classification (function defects, interface problems, performance problems, etc.), problem tags, context of words, and conversations.

[0051] Preprocess the review data. Specifically: clean the review data to remove irrelevant information, including meaningless symbols such as website addresses, special symbols, non-standard emoticons, and common stop words like "of", "is", "in" that are not very helpful for analysis. By cleaning the review data, the overall data quality can be improved, making the data more suitable for machine learning and statistical analysis, speeding up the processing speed, and reducing storage and computing costs.

[0052] Cut the cleaned review data into words with clear meanings to obtain the word segmentation result, which is the preprocessed user review data. For convenience in subsequent analysis and modeling, the word segmentation result can be constructed into a word segmentation list and integrated into structured data.

[0053] S2: Calculate the importance scores of all keywords in the user review data.

[0054] Perform word frequency statistics on the word segmentation in the review data, set a word frequency threshold, and take the words exceeding the threshold as keywords.

[0055] Set the co-occurrence window size, which can be set to 5 - 10 words. Record the co-occurrence times of each keyword with other keywords within the co-occurrence window; calculate the degree centrality of the keywords based on the co-occurrence times of the keywords.

[0056]

[0057] Among them, D k represents the degree centrality of keyword k, and C ki represents the co-occurrence times of keyword k with the i-th keyword.

[0058] Allocate the initial weights of the keywords according to the degree centrality of the keywords. Specifically, the formula for calculating the initial weight value of this keyword is:

[0059] W initial (k) = (D k - D min ) / (D max - D min )

[0060] Among them, W initial (k) represents the initial weight of keyword k, D max represents the maximum degree centrality of the keyword, and D min represents the minimum degree centrality of the keyword.

[0061] Calculate the average position of each keyword in the comments and the average position of all keywords; adjust the initial weight of the keyword according to the average position of the keyword in the comments and the average position of all keywords to obtain the final weight of the keyword. The calculation formula is:

[0062]

[0063] Among them, W final (k) represents the final weight of keyword k, p k represents the average position of keyword k, μ represents the average position of all keywords, and δ represents the standard deviation of keyword positions.

[0064] The final weights of the keywords are normalized to obtain the importance scores of the keywords.

[0065] S3: Select n keywords with the highest importance scores, and construct multi-dimensional features of the comments based on the n keywords and user comment data.

[0066] S31: Calculate the semantic similarity between each keyword and other keywords respectively; take the average of all semantic similarities to obtain the semantic similarity score of the keyword.

[0067] Select n keywords with the highest importance scores, where n is reasonably selected based on the amount of user comment data; calculate the semantic similarity between each keyword and other keywords, and average all the calculated context similarities to obtain the semantic similarity score of the keyword. Preferably, the semantic similarity can be calculated using the cosine similarity between the keyword and other keywords.

[0068] S32: Obtain the contexts of all keywords, and calculate the context similarity between the context of each keyword and the contexts of other keywords respectively; take the average of all context similarities to obtain the context similarity score of the keyword.

[0069] Get the context of all keywords. The context refers to a certain number of words before and after the keywords appear in the comments. Preferably, the window size is set to 7. Calculate the context similarity between the context of each keyword and the context of other keywords. Preferably, the context similarity is the Jaccard similarity, that is, (A∩B) / (A∪B), where A is the context of the keyword and B is the context of another keyword. Take the average of all the calculated context similarities to get the context similarity score of the keyword.

[0070] S33: Perform weighted summation of the semantic similarity score and the context similarity score of the keyword to obtain a comprehensive score of the keyword.

[0071] The weights are designed based on whether the developer cares more about user semantics or context.

[0072] S34: Expand or reduce the context of the keyword according to the comprehensive score of the keyword, and use the context of the keyword as the text feature of the comment to which it belongs.

[0073] Set a comprehensive score threshold, for example, the threshold is 0.5; if the comprehensive score of the keyword is greater than the comprehensive score threshold, expand the context range of the keyword by 2 words before and after, otherwise, narrow the context range of the keyword by 2 words before and after; repeat the above steps, repeatedly calculate the comprehensive score of the keyword, and expand or narrow the context of the keyword until the context range reaches the maximum value or the comprehensive score of the keyword is stable.

[0074] Preferably, the maximum value of the range is set according to the specific application scenario and the characteristics of the data set, and the change range of the comprehensive score within a certain range indicates stability, for example, the score fluctuates within ±0.01.

[0075] S35: Obtain the publishing time of the comment containing the keyword and use it as a time feature.

[0076] S36: Calculate the user score of the comment according to the community feedback of the comment containing the keyword, input the comment containing the keyword into the pre-trained LSTM model to obtain the sentiment score of the comment; obtain the length of the comment containing the keyword.

[0077] Assign weights to the three indicators, namely the number of likes, the number of citations, and the number of replies; perform weighted summation based on the number of likes, the number of citations, and the number of replies of the comment and the weights of their corresponding indicators to obtain the initial user score; normalize the initial user score to obtain the final user score.

[0078] The pre-trained LSTM model is a trained emotion-specific LSTM analysis model that can handle complex emotion expressions that rely on long-distance contexts, such as irony, puns, or emotional changes. It can also capture more complex and subtle emotional details, such as frustration, excitement, and expectations. This is better than simple models that only make simple positive and negative judgments. In the process of continuous learning, the model will gradually forget unimportant information and accept new useful information. This is very effective for dynamically changing data streams (such as real-time user comments). In personalized application scenarios, the model can predict possible future feedback based on the emotional trends of users' historical comments, providing companies with customized customer service strategies.

[0079] The length of the comment containing the keyword, that is, the number of words remaining after removing stop words from the comment.

[0080] S37: Combine the text features, time features, comment length, user scores and sentiment scores of the comments to obtain multi-dimensional features.

[0081] After all the extracted features are preprocessed (missing value filling, outlier processing, etc.), they are standardized to ensure that these different features are on the same scale, and then the features are merged to obtain multi-dimensional features.

[0082] S4: Use the trained adaptive multi-dimensional feature dynamic network to process the multi-dimensional features and obtain the APP software defect recognition results.

[0083] The Adaptive Multidimensional Feature Dynamic Network (AMFDN) includes input layer, feature transformation layer, feature interaction layer, dynamic network layer and output layer.

[0084] The input layer receives multi-dimensional features. The feature transformation layer uses a fully connected layer to perform linear transformation and nonlinear activation on the features. Its structure is:

[0085] Fully connected layer 1: input dimension = total dimension of multi-dimensional features, output dimension = 128

[0086] Activation function: ReLU

[0087] Fully connected layer 2: input dimension = 128, output dimension = 64

[0088] Activation function: ReLU

[0089] The feature interaction layer is used to capture the relationship between different features and uses the attention mechanism to achieve interaction between features. Its structure is as follows:

[0090] Attention layer 1: input dimension = 64, output dimension = 32

[0091] Activation function: ReLU

[0092] Attention layer 2: input dimension = 32, output dimension = 16

[0093] Activation function: ReLU

[0094] The dynamic network layer can automatically increase or decrease the number of network layers according to the feedback of model performance during training; dynamically adjust the number of neurons in each layer according to the complexity and importance of the features; its structure is:

[0095] Dynamic layer 1: input dimension = 16, initial output dimension = 32

[0096] During training, the output dimension is adjusted according to the model performance.

[0097] Dynamic layer 2: input dimension = previous layer output dimension, initial output dimension = 32

[0098] During training, the output dimension is adjusted according to the model performance.

[0099] Dynamic layer 3: input dimension = previous layer output dimension, initial output dimension = 32

[0100] During training, the output dimension is adjusted according to the model performance.

[0101] The output layer is a multi-task learning model that outputs APP software defect identification results.

[0102] Among them, the multi-task learning model can generate a comprehensive output result containing multiple information (such as defect category, defect description, defect severity, sentiment analysis results, and time and frequency analysis). Each task can be trained separately, but share a feature extraction layer so that the model can capture the comprehensive information in the comments. The multi-task learning model can handle multiple tasks simultaneously within a unified framework. This model can share the previous feature extraction layer and then design different output layers for each task. The output layer of each task is defined separately in the model architecture, and the loss of each task is comprehensively calculated during training for optimization. This ensures that the model can output defect category, defect description, defect severity, sentiment analysis results, and time and frequency analysis results at the same time.

[0103] The results of APP software defect identification are:

[0104] 1. Each review is classified into a specific defect category, such as functional defects, performance issues, interface issues, security vulnerabilities, etc. The model outputs the probability distribution of defect categories and selects the category with the highest probability as the identified defect type.

[0105] 2. The model extracts and generates a description of the defect, detailing the specific circumstances of the defect. The description may include key information from user comments, combined with contextual information to form a complete description of the defect.

[0106] 3. The identification results include the severity of the defects, such as high, medium, and low. This can help the development team prioritize severe defects.

[0107] 4. Combined with sentiment analysis, output the user's emotional tendency towards defects (positive, negative, neutral) to help understand the emotions of user feedback.

[0108] 5. The output includes information about the time and frequency of defect occurrence, such as the number of times a defect was reported in a specific time period, which helps identify trends and hot spots.

[0109] For example:

[0110] Suppose a user comments: "The application crashes frequently during operation, affecting the user experience. I hope it will be fixed as soon as possible."

[0111] Identification results will include:

[0112] 1. Defect category: functional defect

[0113] 2. Defect description: "The application frequently crashes during operation"

[0114] 3. Defect severity: High

[0115] 4. Sentiment analysis results: Negative

[0116] 5. Time and frequency analysis: This defect has been reported 15 times in the past week.

[0117] like Figure 2 As shown, during the adaptive multi-dimensional feature dynamic network training process, the MDEM evaluation mechanism is used to verify the network training effect and adjust the network parameters according to the feedback results of the MDEM evaluation mechanism;

[0118] The MDEM evaluation mechanism combines multi-dimensional evaluations, such as accuracy, generalization ability, and prediction consistency, to comprehensively evaluate the model; cross-validation techniques are used to evaluate the stability and effectiveness of the model on different data subsets, and the model parameters are adjusted according to the feedback results of MDEM to optimize the overall performance of the model.

[0119] Preferably, the formula of the MDEM evaluation mechanism is:

[0120]

[0121] Among them, S is the comprehensive score of the model, which is used to evaluate the comprehensive performance of the model; N is the number of test sets, Accuracy i is the accuracy on the i-th test set; K is the number of cross-validation folds, F1-Score k,cross-val is the F1 score of the k-fold cross validation; Var(Predictions) and Var(Inputs) are the variance of the prediction results and the variance of the input data respectively; M is the number of different conditions; Error j,stability is the difference in model stability under the jth condition, where different conditions are different data sets, hyperparameters or network structures; J is the number of consistency tests, Consistency j is the consistency score of the jth test; L is the number of model complexity factors considered, Complexity l is the measure of the lth complexity factor; α, β, γ, δ,∈, ζ are weight factors used to adjust the influence of different dimensions.

[0122] During training, the adaptive multi-dimensional feature dynamic network automatically increases or decreases the number of dynamic network layers or adjusts the number of neurons in each layer according to the training effect to automatically optimize the network structure; preferably, the formula for optimizing the network structure is:

[0123]

[0124] Among them, Θ(l t , E) represents the optimized network structure. Here Θ represents the optimization of the current hierarchical structure l according to the training effect E. t The network hierarchy after adjustment. t represents the current hierarchical structure, and E represents the overall training effect of the adaptive multi-dimensional feature dynamic network, which can be obtained by To calculate, T is the time window size considered, e t represents the training error at time point t, v t represents the validation error at time point t, l t-i represents the current i-th hierarchy, N is the number of historical layers considered; φ, ω, ξ, η are adjustment coefficients.

[0125] During training, the learning rate is dynamically adjusted according to the training error, validation error and the speed of parameter change; preferably, the dynamic learning rate adjustment formula is:

[0126]

[0127] Among them, Λ(l t ,e t ,v t ,ΔP t ) indicates that q t represents the learning rate at time point t; e t represents the training error at time point t; v t represents the verification error at time point t; ΔP t represents the rate of change of model parameters between time t and t-1; α′, β′, δ′, ∈′, γ′ are adjustment coefficients, which can be adjusted according to the specific circumstances of model training.

[0128] By using the trained adaptive multidimensional feature dynamic network to process the multidimensional features, the APP software defect identification results can be obtained. Through the adaptive multidimensional feature dynamic network (AMFDN) model, the present invention can continuously update and adjust the learning model as new data is added. This means that the system can continuously improve itself over time, adapt to new user behaviors and software updates, and thus maintain efficient problem identification capabilities.

[0129] In summary, the present invention uses text analysis technology to process user comments, extract key information, and uses deep learning models to learn and predict potential problems from these data; the present invention identifies software defects by automatically analyzing user comments, thereby greatly reducing the related costs of manual review and greatly improving the efficiency of processing large amounts of data; the present invention can not only improve the efficiency of identifying software problems, but also capture subtle defects that may be missed by traditional testing, thereby providing software developers with a more comprehensive quality assurance tool.

[0130] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying APP software defects based on user reviews, characterized in that: include: S1: Obtain user review data on the software and preprocess it to obtain preprocessed user review data; S2: Calculate the importance scores of all keywords in the user review data; S3: Select n keywords with the highest importance scores, and construct multi-dimensional features of the comments based on the n keywords and user comment data; S4: Use the trained adaptive multi-dimensional feature dynamic network to process the multi-dimensional features and obtain the APP software defect recognition results.

2. The method for identifying APP software defects based on user comments according to claim 1 is characterized in that: The process of preprocessing comment data includes: The comment data is cleaned to remove irrelevant information to obtain the cleaned comment data; the cleaned comment data is cut into words with clear meanings to obtain the word segmentation result, that is, the pre-processed user comment data.

3. The method for identifying APP software defects based on user comments according to claim 1 is characterized in that: The process of calculating a keyword’s importance score involves: Perform word frequency statistics on the words in the comment data, set a word frequency threshold, and use words that exceed the threshold as keywords; Set the co-occurrence window size, record the number of co-occurrences of each keyword with other keywords in the co-occurrence window; calculate the degree centrality of keywords based on the number of co-occurrences of keywords; Assign initial weights to keywords based on their degree centrality; Calculate the average position of each keyword in the comments and the average position of all keywords; adjust the initial weight of the keyword according to the average position of the keyword in the comments and the average position of all keywords to obtain the final weight of the keyword; The final weights of the keywords are normalized to obtain the importance scores of the keywords.

4. The method for identifying APP software defects based on user comments according to claim 3 is characterized in that: The formula for adjusting the initial weight of keywords is: Among them, W final (k) represents the final weight of keyword k, W initial (k) represents the initial weight of keyword k, p k represents the average position of keyword k, μ represents the average position of all keywords, and δ represents the standard deviation of keyword positions.

5. The method for identifying APP software defects based on user comments according to claim 1 is characterized in that: The process of constructing multi-dimensional features includes: S31: Calculate the semantic similarity between each keyword and other keywords respectively; take the average of all semantic similarities to obtain the semantic similarity score of the keyword; S32: Obtain the context of all keywords, and calculate the context similarity between the context of each keyword and the context of other keywords respectively; take the average of all context similarities to obtain the context similarity score of the keyword; S33: performing weighted summation of the semantic similarity score and the context similarity score of the keyword to obtain a comprehensive score of the keyword; S34: expanding or reducing the context of the keyword according to the comprehensive score of the keyword, and using the context of the keyword as the text feature of the comment in which it is located; S35: Obtain the publishing time of the comment containing the keyword and use it as a time feature; S36: Calculate the user score of the comment according to the community feedback of the comment where the keyword is located, input the comment where the keyword is located into the pre-trained LSTM model to obtain the sentiment score of the comment; obtain the length of the comment where the keyword is located; S37: Combine the text features, time features, comment length, user scores and sentiment scores of the comments to obtain multi-dimensional features.

6. The method for identifying APP software defects based on user comments according to claim 5 is characterized in that: The process of expanding or reducing the context of a keyword includes: setting a comprehensive score threshold; if the comprehensive score of the keyword is greater than the comprehensive score threshold, expanding the context range of the keyword by 2 words before and after, otherwise, reducing the context range of the keyword by 2 words before and after; repeatedly calculating the comprehensive score of the keyword, and expanding or reducing the context of the keyword until the context range reaches a maximum value or the comprehensive score of the keyword is stable.

7. The method for identifying APP software defects based on user comments according to claim 5 is characterized in that: The process of calculating the user score of a review includes: Assign weights to the three indicators, namely the number of likes, the number of citations, and the number of replies; perform weighted summation based on the number of likes, the number of citations, and the number of replies of the comment and the weights of their corresponding indicators to obtain the initial user score; normalize the initial user score to obtain the final user score.

8. The method for identifying APP software defects based on user comments according to claim 1, characterized in that: During the training process of the adaptive multi-dimensional feature dynamic network, the MDEM evaluation mechanism is used to verify the network training effect and adjust the network parameters according to the feedback results of the MDEM evaluation mechanism; during training, the adaptive multi-dimensional feature dynamic network optimizes the network structure according to the training effect, and dynamically adjusts the learning rate according to the training error, verification error and the speed of parameter change.

9. The method for identifying APP software defects based on user comments according to claim 8, characterized in that: The formula of the MDEM evaluation mechanism is: Among them, S represents the comprehensive score of the model, N represents the number of test sets, and Accuracy i represents the accuracy on the i-th test set; K represents the number of cross-validation folds, F1-Score k,cross-val represents the F1 score of the k-fold cross validation; Var(Predictions) and Var(Inputs) represent the variance of the prediction results and the variance of the input data respectively; M represents the number of different conditions; Error j,stability represents the model stability difference under the jth condition; J is the number of consistency tests, Consistency j represents the consistency score of the jth test; L is the number of model complexity factors considered, Complexity l represents the measure of the lth complexity factor; α, β, γ, δ,∈, ζ represent the first to sixth weight factors respectively.

10. The method for identifying APP software defects based on user comments according to claim 8, characterized in that: The formula for optimizing the network structure based on the training effect is: Among them, Θ(l t ,E) represents the optimized network structure, l t represents the current hierarchical structure, E represents the overall training effect of the adaptive multi-dimensional feature dynamic network, T represents the time window size considered, and e t represents the training error at time point t, v t represents the validation error at time point t, l t-i represents the current i-th hierarchy, N represents the number of historical layers considered; φ, ω, ξ, η represent the first to fourth adjustment coefficients respectively.

Citation Information

Patent Citations

  • Method and system for analyzing user emotional tendency based on product comments

    CN111695017A

  • Water army identification method and device, electronic equipment and storage medium

    CN115905548A

  • Social network comment hot event public opinion big data analysis system

    CN116361472A

  • Web text public emotion mining method and system based on supervised learning

    CN117076656A

Cited By

  • Natural language recognition processing method and device based on deep learning

    CN120297288A