Player feedback keyword extraction method based on big data model

Through the big data model and BERT+TextRank method, combined with the multi-task learning framework, the problem of extracting feedback from players on multiple platforms is solved, and accurate analysis of off-game feedback and rapid optimization of game design is achieved.

CN118917309APending Publication Date: 2024-11-08ZHEJIANG UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410944839.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing technology cannot effectively obtain and analyze players' feedback from multiple platforms outside the game, resulting in the inability to follow up on external game players' feedback and extract effective game key information in a timely manner.

Method used

Using a player feedback keyword extraction method based on big data model, collect and preprocess forum and social media data through Python crawlers, extract keywords using BERT model and TextRank algorithm, and integrate multi-task learning framework for sentiment analysis and domain correlation analysis and behavior-oriented analysis, dynamically adjust the judgment threshold of extracting keywords, and generate iterative solutions to optimize game design.

Benefits of technology

It improves the accuracy of keyword extraction of multi-platform data, helps developers quickly locate player feedback problems and optimize game design, ensuring the accuracy of keyword extraction and system responsiveness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917309A_ABST
    Figure CN118917309A_ABST
Patent Text Reader

Abstract

The invention discloses a player feedback keyword extraction method based on a big data model. The method comprises the steps that a Python crawler is used for obtaining data related to a game and preprocessing the data; training a similarity measurement standard related to the game, and establishing a corresponding task window; carrying out keyword extraction on the text data in different task windows by using BERT and TextRank; an MTL framework is integrated, multi-task optimization including sentiment analysis, field correlation analysis and behavior oriented analysis is carried out on the basis of the extracted keywords, and final keywords are output; the output final keyword is fed back to a multi-task learning MTL model, and a judgment threshold value of the extracted keyword is dynamically adjusted; and bringing the obtained keyword data into an analysis model, feeding back to generate an optimal game iteration scheme, and updating and optimizing the game. By means of the method and device, feedback of players outside the game can be followed in time, the accuracy of effective keyword extraction of multi-platform information is improved, and therefore developers are helped to rapidly position the problems fed back by the players and optimize game design.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of keyword extraction, and in particular relates to a player feedback keyword extraction method based on a big data model. Background Art

[0002] In today's gaming industry, understanding and analyzing player feedback has become one of the core competitive advantages. Game developers and operators are constantly seeking to optimize the gaming experience and improve user satisfaction by analyzing player discussions on forums, social media, and review platforms. To this end, a player feedback keyword extraction system based on big data models is particularly important. Such a system can not only help game companies obtain and analyze player feedback in real time, but also generate targeted improvement measures based on this, so as to quickly respond to market changes and player needs. Keyword extraction is to identify key areas and issues that players are concerned about by analyzing the frequent occurrence and key words in the player's language. By identifying keywords, the development team can quickly locate the problem and adjust or optimize the game design and server configuration accordingly to improve the player's gaming experience.

[0003] Existing keyword extraction methods, such as the keyword extraction method, device, equipment and storage medium of the text with publication number CN114020876A, determine the target keyword in the text to be analyzed through the ambiguous prior probability of the initial keyword and the preset ambiguous prior model, so that the accuracy of the target keyword in the text to be analyzed is improved. Another example is the player-posted content processing method and device with publication number CN106709054A, which obtains the content posted by the player on the game information publishing platform through a crawler, extracts the corresponding game follow-up personnel for keyword matching, and promptly answers questions or feedbacks to technical personnel on the platform for processing, thereby providing the possibility of improving the game experience.

[0004] However, the above-mentioned existing keyword extraction methods do not involve user feedback outside the game itself and the official game platform. Now the official feedback channels for players are mainly official game forums or questionnaires, and the other part is highly relevant to the game content and has real-time comments and opinions, which are scattered on other social platforms and comprehensive forums. This part of content, as direct feedback from players, is important information for the game to improve user experience. In addition, the above-mentioned existing keyword extraction methods do not involve the extraction of effective key information on game content after obtaining multi-platform information. For the problem that the existing methods cannot follow up on player feedback outside the game in a timely manner, and how to extract effective key game information after obtaining complex information from various platforms, no effective solution has been proposed yet. Summary of the invention

[0005] The present invention provides a player feedback keyword extraction method based on a big data model, which can improve the accuracy of keyword extraction from multi-platform data and generate suggestions and iterative solutions, thereby helping developers quickly locate problems reported by players and optimize game design.

[0006] A method for extracting player feedback keywords based on a big data model comprises the following steps:

[0007] (1) Use Python to crawl forums and social media to collect and preprocess game-related player feedback and post data;

[0008] (2) Based on the preprocessed data, the similarity measurement criteria related to the training game are established to establish the corresponding task window;

[0009] (3) Use the BERT model and TextRank algorithm to extract keywords from text data in different task windows;

[0010] (4) Integrate the MTL framework and perform multi-task optimization based on the keywords extracted in step (3), including sentiment analysis, domain relevance analysis, and behavior orientation analysis, and output the final keywords;

[0011] (5) Feed the final keywords output to the multi-task learning MTL model and dynamically adjust the judgment threshold of keyword extraction;

[0012] (6) Based on the continuously obtained keyword data, the keyword data is incorporated into the analysis model, and feedback is used to generate the optimal game iteration plan to update and optimize the game content.

[0013] The specific process of step (1) is as follows:

[0014] (1-1) Define the Spider, set the starting URL and parsing function to crawl and parse forum and social media data;

[0015] (1-2) Run the crawler and collect data;

[0016] (1-3) Clean and purify the collected text data, remove duplicates, unify the data format, conduct preliminary analysis of the data, and correct or delete outliers;

[0017] (1-4) The captured data is stored in a database for subsequent analysis.

[0018] The specific process of step (2) is as follows:

[0019] (2-1) Based on the preprocessed data, multiple task windows are defined according to different content areas of the game, including system performance feedback, player interaction experience, and game content feedback;

[0020] (2-2) Customize processing rules for each task window, including identification and classification of specific terms, to ensure that the data is highly relevant to the task;

[0021] (2-3) Use the game-specific pre-trained BERT model to extract semantic embedding vectors for each task window;

[0022] (2-4) Train a similarity calculation model for each task window, and use different machine learning algorithms and parameter settings according to different window characteristics to ensure that the model accurately captures the content characteristics of a specific window;

[0023] (2-5) Dynamically adjust the similarity threshold according to the characteristics and feedback of each task window to optimize the accuracy of keyword extraction and content analysis;

[0024] (2-6) Integrate the multi-task learning (MTL) framework, feedback the output of each task window and user feedback, and adjust the processing flow.

[0025] The specific process of step (3) is as follows:

[0026] (3-1) Using BERT, the preprocessed text words are deeply semantically embedded to generate high-dimensional vector representations. These embedded vectors capture the semantic information of each word in its context;

[0027] (3-2) Using the embedding vectors generated by BERT, we calculate the similarity between words in the text, measured by the cosine similarity between word vectors, and build a semantic similarity graph based on these similarity scores, where the nodes represent words and the edge weights reflect the semantic similarity between words;

[0028] (3-3) Apply the TextRank algorithm on the constructed semantic similarity graph and iteratively calculate the score of each node until convergence; sort the words according to the TextRank score and select the words whose scores meet the requirements as keywords.

[0029] In step (3-1), the preprocessed text vocabulary is deeply semantically embedded to generate a high-dimensional vector representation. The specific process is as follows:

[0030] (3-1-1) Use BERT’s tokenizer to decompose the text into words or subword units;

[0031] (3-1-2) The segmented text is converted into an input format that the model can understand, including word IDs, segmentation tags, and attention masks;

[0032] (3-1-3) Input the text processed in step (3-1-2) into the BERT model. The BERT model processes the input through its multi-layer Transformer network. Each layer of Transformer contains a self-attention mechanism and a feedforward network, so that the representation of each word can integrate the information of surrounding words;

[0033] (3-1-4) The BERT model outputs embedding vectors for each word. These vectors are high-dimensional and can richly represent the semantic features of each word.

[0034] The specific process of step (3-2) is:

[0035] (3-2-1) Use the BERT model to generate an embedding vector for each word in the text;

[0036] (3-2-2) The cosine similarity is defined by calculating the ratio of the dot product of two vectors to the product of their moduli. The formula is:

[0037]

[0038] Where v1 and v2 are the embedding vectors of the two words, the dot product v1·v2 represents the dot product between the vectors, and ||v1|| and ||v2|| are the Euclidean norms of the two vectors respectively;

[0039] (3-2-3) The initial threshold is set to 0.7. If the cosine similarity is greater than or equal to the threshold, the two word vectors are considered to be similar in content. Otherwise, they are considered to be dissimilar.

[0040] (3-2-4) When the similarity exceeds the threshold, an edge is established between two nodes, and the final result is a semantic similarity graph;

[0041] The machine learning model adjusts the threshold value range according to the keyword extraction effect in the actual feedback, and predicts and uses the optimal threshold.

[0042] The specific process of step (3-3) is:

[0043] (3-3-1) The TextRank algorithm is applied to iteratively calculate the score of each node. The core formula of TextRank is as follows:

[0044]

[0045] Among them, TR(v i ) is the node V i score; d is the damping factor, set to 0.8; V j It points to V i The node set I n (Vi ) in the node; ω ji is from node V j To V i The weight of the edge; ω jk is from node V j The formula iteratively updates the importance score of each node based on the amount of support each node receives.

[0046] (3-3-2) After the calculation is completed, all nodes are sorted according to the scores to obtain a score list sorted from low to high, and the 80th percentile value words and subsequent words in all words are selected as keywords.

[0047] In step (4), the MTL framework is integrated to share the underlying representation of BERT. While running the BERT and TextRank tasks, the sentiment analysis, domain relevance, and behavior-oriented word tasks are maintained to optimize the keyword extraction effect.

[0048] In step (4), the specific process of sentiment analysis, domain relevance analysis and behavior orientation analysis is as follows:

[0049] (4-1) Define sentiment carrier vocabulary and analyze the sentiment tendency of keywords;

[0050] (4-2) Define keyword domain relevance, especially vocabulary specific to the gaming domain;

[0051] (4-3) Define action-oriented vocabulary, including vocabulary that may trigger player actions or requests;

[0052] (4-4) Analyze keyword features, improve data-driven decision-making, and enhance the accuracy and processing precision of the system’s understanding of text.

[0053] The specific process of step (6) is as follows:

[0054] (6-1) Retrieving the obtained keyword data, classifying and weighting the keywords, and identifying which keywords are most relevant to player satisfaction, game criticism, and specific needs;

[0055] (6-2) Using the extracted keywords and player behavior data, player feedback data, player interaction data, system performance log data, and error report data, construct an analysis and prediction model to predict the impact of content or issues corresponding to specific keywords on player retention and satisfaction, and use machine learning algorithms to analyze the impact of key content on player behavior;

[0056] (6-3) Based on the analysis results of the model, generate strategy texts and iterative plans for improving the game, including server latency, character customization requirements, and team copy difficulty adjustments;

[0057] (6-4) Continuously monitor keywords and user feedback, and constantly adjust analysis models and game content. The system dynamically adjusts strategies through continuous data analysis to respond to new player needs.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1. The present invention starts from the perspective of keyword features, classifies keywords through domain relevance, sentiment analysis and behavior-oriented analysis, and captures the deep semantic relationship and contextual information of vocabulary through the BERT+TextRank method, taking into account both word frequency and word meaning, so as to extract truly important keywords.

[0060] 2. The present invention integrates keyword output and supports analysis result feedback to dynamically adjust the judgment threshold for extracting keywords, including the weight of judging keywords and the threshold of cosine similarity of semantic similarity graph, so as to improve the accuracy of system recognition and processing of keywords.

[0061] 3. The present invention supplements and extracts highly relevant feedback data of games on other platforms. The background database can update information in real time, generate reports based on the analysis results, and send diagnostic suggestions to help developers quickly locate problems and optimize game design. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a method for extracting player feedback keywords based on a big data model according to an embodiment of the present invention;

[0063] Figure 2 A schematic diagram of a flow chart of text preprocessing for converting text words into high-dimensional vectors in an embodiment of the present invention;

[0064] Figure 3 A schematic diagram of a process for generating a high-dimensional vector for a vocabulary in an embodiment of the present invention;

[0065] Figure 4 The figure is a flow chart of extracting keywords using the BERT model and the TextRank algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0067] like Figure 1 As shown, a player feedback keyword extraction method based on a big data model includes the following steps:

[0068] S1. Define the Spider, set the starting URL and parsing function to crawl and parse web post data, and use Python to crawl related forums and social media for data collection.

[0069] S2. Based on the collected data, pre-process the data, including cleaning and purifying the text data, deduplicating, unifying the data format, and performing preliminary analysis on the data to identify and process outliers for correction or deletion.

[0070] S3. Based on the preprocessed data, the similarity measurement standard related to the training game is established to establish the corresponding task window.

[0071] S301. Based on the preprocessed data, define multiple task windows according to different content areas of the game, such as "system performance feedback", "player interactive experience" and "game content feedback".

[0072] S302, customize preprocessing rules for each task window, including identification and classification of specific terms, to ensure that the data is highly relevant to the task. For example, in the system performance feedback window, identify and classify feedback about game performance, such as loading time, delay, server errors, etc. The preprocessing rules for system performance feedback are to first identify and extract professional terms about performance, such as "lag", "crash", "frame rate", etc., and then remove discussions that are not related to performance, such as game strategy and character development, and finally convert synonyms into standard terms, such as unifying "slow response" into "lag".

[0073] S303: Use the game-specific pre-trained BERT model to extract a semantic embedding vector for each task window.

[0074] S304, train the similarity calculation model for each task window, and use different machine learning algorithms and parameter settings according to different window features to ensure that the model accurately captures the content characteristics of a specific window. For example, in the sentiment analysis task window, implement a fine-grained sentiment classifier based on BERT to evaluate sentiment tendencies (positive, negative, neutral), and in the domain relevance task window, extract the preprocessed text high-dimensional vector and input it into the topic model, use the clustering algorithm to obtain the topic, and use it in subsequent content analysis and decision support.

[0075] S305. Dynamically adjust the similarity threshold according to the characteristics and feedback of each task window to optimize the accuracy of keyword extraction and content analysis.

[0076] S306, integrating a multi-task learning (MTL) framework, feedback looping the output of each task window and user feedback, and adjusting the processing flow.

[0077] S4. Use the BERT model and TextRank algorithm to extract keywords from text data in different task windows.

[0078] S401, deep semantic embedding of vocabulary is generated to generate high-dimensional vector representations, which capture the semantic information of each word in its context. Before converting to high-dimensional vectors, the text needs to be preprocessed, such as Figure 2 As shown, the preprocessing process is:

[0079] 1. Preprocess the text converted into high-dimensional vectors;

[0080] 2. Use the WordPiece algorithm for word segmentation. BERT's word segmenter breaks down the text into vocabulary units (tokens) and handles unknown or rare words;

[0081] 3. Add special tags `[CLS]` and `[SEP]` at the beginning and end of the sentence respectively. The `[CLS]` tag is used for the output of certain specific tasks (such as classification tasks), and the `[SEP]` is used to separate sentences or different text paragraphs;

[0082] 4. Generate an attention mask to allow the model to distinguish between the content that should be paid attention to and the padding part, and handle inputs of different lengths.

[0083] 5. Output data is used to convert into model input.

[0084] like Figure 3 As shown, the process of generating high-dimensional vectors from vocabulary in the embodiment of the present invention is as follows:

[0085] 1. Convert each token to the corresponding index ID in the pre-trained vocabulary;

[0086] 2. According to the maximum length limit of the BERT model (for example, 512 tokens), truncate the input that is too long and pad the input that is too short;

[0087] 3. Generate positional encoding and use the BERT model to use positional encoding to understand the position of words in a sentence to understand the order information in the language;

[0088] 4. Input the BERT model and input the processed token ID, attention mask and position encoding into the BERT model;

[0089] 5. Through the Transformer layer, each layer includes a self-attention mechanism and a feed-forward neural network to process the input and refine the information layer by layer;

[0090] 6. The model outputs embedding vectors for each token, which are high-dimensional (for example, the vector dimension generated by the BERT base model is 768);

[0091] 7. Select different output layers to obtain word vectors as needed. For specific tokens (such as those used for named entity recognition or keyword extraction), directly extract the vectors corresponding to these tokens from the output.

[0092] S402: Use the BERT model and TextRank algorithm to extract keywords. The process is as follows: Figure 4 shown.

[0093] Use the BERT model to generate an embedding vector for each word in the text; calculate the ratio of the dot product of two vectors to the product of their modulos to define the cosine similarity, the formula is:

[0094]

[0095] Where v1 and v2 are the embedding vectors of the two words, the dot product v1·v2 represents the dot product between the vectors, and ||v1|| and ||v2|| are the Euclidean norms of the two vectors respectively;

[0096] The initial threshold is set to 0.7. If the cosine similarity is greater than or equal to the threshold, the two word vectors are considered to be similar in content. Otherwise, they are considered to be dissimilar.

[0097] When the similarity exceeds the threshold, an edge is established between the two nodes, and the final result is a semantic similarity graph;

[0098] The TextRank algorithm is applied to iteratively calculate the score of each node. The core formula of TextRank is as follows:

[0099]

[0100] Among them, TR(v i ) is the node V j score; d is the damping factor, set to 0.8; V j It points to V i The node set I n (V i ) in the node; ω ji is from node V j To V i The weight of the edge; ω jk is from node V j The formula iteratively updates the importance score of each node based on the amount of support each node receives.

[0101] After the calculation is completed, all nodes are sorted according to the scores to obtain a score list sorted from low to high, and the 80th percentile value words and subsequent words in all words are selected as keywords.

[0102] S403. Share the underlying representation of BERT. While running BERT and TextRank tasks, keep other tasks (sentiment analysis, domain relevance, and behavior-oriented words) running to optimize keyword extraction effects.

[0103] S404. Refine and optimize the parameters of the model through feedback from actual applications, including adjusting the threshold of TextRank or improving the fine-tuning process of BERT to better adapt to specific text data and needs.

[0104] S5. Integrate the MTL framework and perform multi-task optimization based on the initially extracted keywords, including sentiment analysis, domain relevance analysis, and behavior-oriented analysis, to output the final keywords.

[0105] The specific process of sentiment analysis, domain relevance analysis and behavior-oriented analysis is as follows:

[0106] Define keyword domain relevance, especially those specific to the gaming field, such as "raid", "guild", "PvP" and other words in MMORPG games;

[0107] Define sentiment carrier words and analyze the sentiment tendency of keywords, such as "amazing", "frustrated", "buggy" and other sentiment-containing words;

[0108] Define action-oriented words, including words that may trigger player actions or requests, such as "need", "wish", and "suggest";

[0109] Analyze keyword features to improve data-driven decision making and increase the accuracy and processing precision of the system’s understanding of text.

[0110] S6. Feed the final output keywords back to the multi-task learning MTL model to dynamically adjust the judgment threshold of keyword extraction.

[0111] S7. Continuously monitor keywords and user feedback, generate diagnostic suggestion reports, locate problems, and continuously adjust analysis models and game content. The system dynamically adjusts strategies and iterates optimization through continuous data analysis to respond to new player needs.

[0112] Based on the same inventive principle, an embodiment of the present invention also provides a player feedback keyword extraction system based on a big data model, including a memory and one or more processors, wherein executable code is stored in the memory, and when one or more processors execute the executable code, it is used to implement the above-mentioned player feedback keyword extraction method.

[0113] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A player feedback keyword extraction method based on a big data model, characterized in that: The following steps are involved: (1) Use Python to crawl forums and social media to collect and preprocess game-related player feedback and post data; (2) Based on the preprocessed data, the similarity measurement criteria related to the training game are established to establish the corresponding task window; (3) Use the BERT model and TextRank algorithm to extract keywords from text data in different task windows; (4) Integrate the MTL framework and perform multi-task optimization based on the keywords extracted in step (3), including sentiment analysis, domain relevance analysis, and behavior orientation analysis, and output the final keywords; (5) Feed the final keywords output to the multi-task learning MTL model and dynamically adjust the judgment threshold of keyword extraction; (6) Based on the continuously obtained keyword data, the keyword data is incorporated into the analysis model, and feedback is used to generate the optimal game iteration plan to update and optimize the game content.

2. The player feedback keyword extraction method based on the big data model according to claim 1 is characterized in that: The specific process of step (1) is as follows: (1-1) Define the Spider, set the starting URL and parsing function to crawl and parse forum and social media data; (1-2) Run the crawler and collect data; (1-3) Clean and purify the collected text data, remove duplicates, unify the data format, conduct preliminary analysis of the data, and correct or delete outliers; (1-4) The captured data is stored in a database for subsequent analysis.

3. The player feedback keyword extraction method based on the big data model according to claim 1 is characterized in that: The specific process of step (2) is as follows: (2-1) Based on the preprocessed data, multiple task windows are defined according to different content areas of the game, including system performance feedback, player interaction experience, and game content feedback; (2-2) Customize processing rules for each task window, including identification and classification of specific terms, to ensure that the data is highly relevant to the task; (2-3) Use the game-specific pre-trained BERT model to extract semantic embedding vectors for each task window; (2-4) Train a similarity calculation model for each task window, and use different machine learning algorithms and parameter settings according to different window characteristics to ensure that the model accurately captures the content characteristics of a specific window; (2-5) Dynamically adjust the similarity threshold according to the characteristics and feedback of each task window to optimize the accuracy of keyword extraction and content analysis; (2-6) Integrate the multi-task learning (MTL) framework, feedback the output of each task window and user feedback, and adjust the processing flow.

4. The player feedback keyword extraction method based on the big data model according to claim 1 is characterized in that: The specific process of step (3) is as follows: (3-1) Using BERT, the preprocessed text words are deeply semantically embedded to generate high-dimensional vector representations. These embedded vectors capture the semantic information of each word in its context; (3-2) Using the embedding vectors generated by BERT, we calculate the similarity between words in the text, measured by the cosine similarity between word vectors, and build a semantic similarity graph based on these similarity scores, where the nodes represent words and the edge weights reflect the semantic similarity between words; (3-3) Apply the TextRank algorithm on the constructed semantic similarity graph and iteratively calculate the score of each node until convergence; sort the words according to the TextRank score and select the words whose scores meet the requirements as keywords.

5. The player feedback keyword extraction method based on the big data model according to claim 4 is characterized in that: In step (3-1), the preprocessed text vocabulary is deeply semantically embedded to generate a high-dimensional vector representation. The specific process is as follows: (3-1-1) Use BERT’s tokenizer to decompose the text into words or subword units; (3-1-2) The segmented text is converted into an input format that the model can understand, including word IDs, segmentation tags, and attention masks; (3-1-3) Input the text processed in step (3-1-2) into the BERT model. The BERT model processes the input through its multi-layer Transformer network. Each layer of Transformer contains a self-attention mechanism and a feedforward network, so that the representation of each word can integrate the information of surrounding words; (3-1-4) The BERT model outputs embedding vectors for each word. These vectors are high-dimensional and can richly represent the semantic features of each word.

6. The player feedback keyword extraction method based on big data model according to claim 4 is characterized in that: The specific process of step (3-2) is: (3-2-1) Use the BERT model to generate an embedding vector for each word in the text; (3-2-2) The cosine similarity is defined by calculating the ratio of the dot product of two vectors to the product of their moduli. The formula is: Where v1 and v2 are the embedding vectors of the two words, the dot product v1·v2 represents the dot product between the vectors, and ||v1|| and ||v2|| are the Euclidean norms of the two vectors respectively; (3-2-3) The initial threshold is set to 0.

7. If the cosine similarity is greater than or equal to the threshold, the two word vectors are considered to be similar in content. Otherwise, they are considered to be dissimilar. (3-2-4) When the similarity exceeds the threshold, an edge is established between two nodes, and the final result is a semantic similarity graph; The machine learning model adjusts the threshold value range according to the keyword extraction effect in the actual feedback, and predicts and uses the optimal threshold.

7. The player feedback keyword extraction method based on big data model according to claim 4 is characterized in that: The specific process of step (3-3) is: (3-3-1) The TextRank algorithm is applied to iteratively calculate the score of each node. The core formula of TextRank is as follows: Among them, TR(v i ) is the node V i score; d is the damping factor, set to 0.8; V j It points to V i The node set I n (V i ) in the node; ω ji is from node V j To V i The weight of the edge; ω jk is from node V j The formula iteratively updates the importance score of each node based on the amount of support each node receives. (3-3-2) After the calculation is completed, all nodes are sorted according to the scores to obtain a score list sorted from low to high, and the 80th percentile value words and subsequent words in all words are selected as keywords.

8. The player feedback keyword extraction method based on big data model according to claim 4 is characterized in that: In step (4), the MTL framework is integrated to share the underlying representation of BERT. While running the BERT and TextRank tasks, the sentiment analysis, domain relevance, and behavior-oriented word tasks are maintained to optimize the keyword extraction effect.

9. The player feedback keyword extraction method based on big data model according to claim 1, characterized in that: In step (4), the specific process of sentiment analysis, domain relevance analysis and behavior orientation analysis is as follows: (1) Define sentiment carrier vocabulary and analyze the sentiment tendency of keywords; (2) define keyword domain relevance, especially vocabulary specific to the gaming domain; (3) define action-oriented words, including words that may trigger player actions or requests; (4) Analyze keyword features, improve data-driven decision-making, and enhance the accuracy of the system’s understanding of text and processing precision.

10. The player feedback keyword extraction method based on big data model according to claim 1, characterized in that: The specific process of step (6) is as follows: (6-1) Retrieving the obtained keyword data, classifying and weighting the keywords, and identifying which keywords are most relevant to player satisfaction, game criticism, and specific needs; (6-2) Using the extracted keywords and player behavior data, player feedback data, player interaction data, system performance log data, and error report data, construct an analysis and prediction model to predict the impact of content or issues corresponding to specific keywords on player retention and satisfaction, and use machine learning algorithms to analyze the impact of key content on player behavior; (6-3) Based on the analysis results of the model, generate strategy texts and iterative plans for improving the game, including server latency, character customization requirements, and team copy difficulty adjustments; (6-4) Continuously monitor keywords and user feedback, and constantly adjust analysis models and game content. The system dynamically adjusts strategies through continuous data analysis to respond to new player needs.

Citation Information

Cited By

  • Hot news prediction scheduling method and system under big data platform

    CN119721410A

  • Multi-platform-based copywriting keyword recognition and analysis method, device and equipment

    CN121638230A