Game community entity word mining method and device, electronic equipment and storage medium

Abstract

The application provides a game community entity word mining method and device, electronic equipment and storage medium. The method comprises the following steps: obtaining text corpus corresponding to a plurality of game communities, and generating a target corpus set; based on the co-occurrence statistical result and the edge probability distribution of adjacent characters in the target corpus set, the cohesion degree corresponding to the character connection position is calculated, and the low cohesion degree segmentation position is identified based on the preset cohesion degree threshold; according to the low cohesion degree segmentation position, reverse segmentation is performed on the target corpus set to obtain a candidate stable segment set; the frequency distribution of each candidate word in the plurality of game communities and the plurality of game identifiers is counted, and the cross-community public word is filtered out from the candidate word set; the entity confirmation and sorting are performed on the filtered candidate word combined with the game context trigger feature and the context co-occurrence relationship, and the entity word mining result corresponding to the target game community is output. The application can reduce the candidate calculation scale, suppress the public word interference, and improve the community entity word mining accuracy.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, electronic device, and storage medium for mining entity words in game communities. Background Technology

[0002] With the continuous development of the online game industry and game community ecosystem, players have generated a large number of posts, comments, bullet comments, and strategy texts related to characters, equipment, dungeons, events, versions, and gameplay strategies. This type of text is frequently updated, colloquial, and contains numerous abbreviations and aliases. Furthermore, the same entity often appears in different communities using abbreviations, colloquialisms, or mixed symbols. Therefore, accurately extracting game entity terms from massive game community corpora has become a fundamental technical challenge in game content understanding, community knowledge graph construction, and entity database maintenance.

[0003] Existing methods for new word discovery or entity mining are mostly geared towards general corpora. They typically employ dictionary-based matching, forward enumeration of candidate strings, and filtering based on frequency, cohesion, and adjacency entropy, or directly utilize general word segmentation models for candidate identification. When these methods are applied to game community scenarios, they often require traversing and calculating a large number of character combinations, resulting in a large processing scale. Furthermore, general methods struggle to distinguish between game community-specific entity words and generalized words that commonly appear across different communities, easily misidentifying common expressions such as "ranking up," "strategy," and "why" as target entity words, leading to insufficient accuracy and community-specificity in entity word results. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, electronic device, and storage medium for mining entity words in game communities, in order to solve the problems of large candidate fragment computation scale, strong interference from common words, and difficulty in identifying community-specific entities in the prior art.

[0005] A first aspect of this application provides a method for mining entity words in game communities, comprising: acquiring text corpora corresponding to multiple game communities; performing noise reduction and normalization, mixed symbol normalization, and association processing of community identifiers and game identifiers on the text corpora to generate a target corpus set; calculating the cohesion corresponding to the character connection position based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus set, and determining a preset cohesion threshold based on the cohesion distribution of the entire corpus, and identifying low-cohesion segmentation positions based on the preset cohesion threshold; performing reverse segmentation on the target corpus set according to the low-cohesion segmentation positions to obtain a candidate stable segment set, and assembling a candidate word set based on the adjacency relationship and length constraint of the candidate stable segment set; statistically analyzing the frequency distribution of each candidate word under multiple game communities and multiple game identifiers, identifying cross-community common words based on community coverage, game coverage, and distribution dispersion, and filtering out cross-community common words from the candidate word set; performing entity confirmation and sorting on the filtered candidate words in combination with game context triggering features and context co-occurrence relationships, and outputting the entity word mining results corresponding to the target game community.

[0006] A second aspect of this application provides a device for mining entity words in game communities, comprising: an acquisition module, configured to acquire text corpora corresponding to multiple game communities, perform noise reduction and normalization, mixed symbol normalization, and association processing of community identifiers and game identifiers on the text corpora to generate a target corpus set; a calculation module, configured to calculate the cohesion corresponding to the character connection position based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus set, determine a preset cohesion threshold based on the cohesion distribution of the entire corpus, and identify low-cohesion segmentation positions based on the preset cohesion threshold; and a segmentation module, configured to segment based on low cohesion... The segmentation module performs reverse segmentation on the target corpus to obtain a set of candidate stable segments, and assembles a set of candidate words based on the adjacency relationship and length constraints of the candidate stable segments. The identification module is used to count the frequency distribution of each candidate word under multiple game communities and multiple game identifiers, identify cross-community common words based on community coverage, game coverage and distribution dispersion, and filter out cross-community common words from the candidate word set. The output module is used to perform entity confirmation and sorting on the filtered candidate words in combination with game context trigger features and context co-occurrence relationship, and output the entity word mining results corresponding to the target game community.

[0007] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0008] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0009] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring text corpora corresponding to multiple game communities, this application performs noise reduction and normalization, mixed symbol normalization, and association processing of community and game identifiers to generate a target corpus. Based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus, the cohesion of character connection positions is calculated, and a preset cohesion threshold is determined according to the cohesion distribution of the entire corpus. Low-cohesion segmentation positions are identified based on the preset cohesion threshold. The target corpus is then reverse-segmented according to the low-cohesion segmentation positions to obtain a set of candidate stable segments. A candidate word set is assembled based on the adjacency relationship and length constraints of the candidate stable segment set. The frequency distribution of each candidate word under multiple game communities and multiple game identifiers is statistically analyzed. Cross-community common words are identified based on community coverage, game coverage, and distribution dispersion, and cross-community common words are filtered out from the candidate word set. Entity confirmation and ranking are performed on the filtered candidate words in combination with game context trigger features and context co-occurrence relationships, and the entity word mining results corresponding to the target game community are output. This application can reduce the candidate computation scale, suppress common word interference, and improve the accuracy of community entity word mining. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating the game community entity mining method provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the game community entity word mining device provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0013] In existing technologies, the mining of new words or entities in text typically employs dictionary matching, general word segmentation models for identification, or forward enumeration of character combinations in the corpus, followed by filtering based on statistical indicators such as frequency, cohesion, and adjacency entropy. These methods are primarily designed for general text scenarios and are applicable to news, web pages, or general social media texts. However, game community corpora are characterized by strong colloquialism, frequent updates, numerous abbreviations and aliases, widespread mixing of Chinese characters with alphanumeric characters, and multiple names for the same entity, making existing general methods insufficiently adaptable to the game community scenario.

[0014] Based on the aforementioned existing technologies, the following problems easily arise during the mining of entity words in game communities: First, directly enumerating and calculating a large number of character combinations leads to a large computational scale in the candidate fragment generation process; second, common words that frequently appear across communities, such as "ranking up," "strategy," and "why," are easily mixed into the candidate results, interfering with community-specific entity words; third, existing methods struggle to effectively confirm candidate words by incorporating the contextual features of game communities, resulting in insufficient accuracy and discriminative power in identifying community entity words. Therefore, how to control the computational scale of candidate words while reducing interference from common words and improving the recognition capability of game community-specific entity words becomes the technical problem that this application aims to solve.

[0015] To address the aforementioned technical issues, this application proposes a method for mining entity words in game communities. First, text corpora corresponding to multiple game communities are acquired. These corpora undergo denoising, normalization, mixed symbol normalization, and association processing between community and game identifiers to generate a target corpus. Then, based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus, the cohesion of character connection positions is calculated. A preset cohesion threshold is determined based on the overall corpus cohesion distribution to identify low-cohesion segmentation positions. Next, the target corpus is reverse-segmented based on these low-cohesion segmentation positions to obtain a set of candidate stable segments. Finally, a candidate word set is assembled based on the adjacency relationships and length constraints of the candidate stable segment set.

[0016] Building upon this foundation, this application further analyzes the frequency distribution of each candidate word across multiple game communities and game identifiers. Based on community coverage, game coverage, and distribution dispersion, it identifies cross-community common words and filters them out from the candidate word set. Subsequently, it performs entity confirmation and ranking on the filtered candidate words, combining game contextual trigger features and contextual co-occurrence relationships, and outputs the entity word mining results corresponding to the target game community. This technical solution does not simply follow the general new word discovery process; instead, it combines low-cohesion reverse segmentation with cross-community common word filtering to form an entity word mining chain suitable for game community scenarios.

[0017] By adopting the above technical solutions, this application can reduce the computational scale of the candidate fragment generation stage, suppress the interference of cross-community public words on entity word mining results, and improve the accuracy and community distinguishability of game community entity word mining, thereby providing relatively stable basic data support for subsequent community entity database construction, knowledge graph organization and game content understanding.

[0018] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] Figure 1 This is a flowchart illustrating the game community entity word mining method provided in this application embodiment. For example... Figure 1 As shown, the specific entity mining method for this game community can include: S101: Obtain text corpora corresponding to multiple game communities, perform noise reduction and normalization, mixed symbol normalization, and association processing of community identifiers and game identifiers on the text corpora to generate the target corpus set; S102, based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus, calculate the cohesion corresponding to the character connection position, and determine the preset cohesion threshold according to the cohesion distribution of the whole corpus, and identify low cohesion segmentation positions based on the preset cohesion threshold. S103, Perform reverse segmentation on the target corpus based on the low cohesion segmentation position to obtain a set of candidate stable segments, and assemble a set of candidate words based on the adjacency relationship and length constraints of the candidate stable segment set; S104, Statistically analyze the frequency distribution of each candidate word in multiple game communities and multiple game identifiers, identify cross-community common words based on community coverage, game coverage and distribution dispersion, and filter out cross-community common words from the candidate word set; S105, after filtering out candidate words, combines game context trigger features and context co-occurrence relationships to perform entity confirmation and ranking, and outputs the entity word mining results corresponding to the target game community.

[0020] In some embodiments, text corpora corresponding to multiple game communities are obtained, and noise reduction, normalization, mixed symbol normalization, and association processing of community identifiers and game identifiers are performed on the text corpora to generate a target corpus set, including: Collect post titles, texts, replies, strategy comments, bullet screen texts, battle report descriptions, and event discussion texts from various game communities, and extract the posting time, community source, section identifier, interaction popularity marker, and game information corresponding to each text. The text is processed by merging duplicate content, cleaning links and emojis, compressing abnormal characters, unifying full-width and half-width characters, converting between traditional and simplified Chinese characters, and removing noisy segments to obtain standardized text segments. The system identifies mixed symbol strings such as version numbers, season numbers, region codes, equipment numbers, and character abbreviations in standardized text fragments, normalizes them according to preset retention rules and alias unification rules, and establishes a hierarchical association index based on community source, section identifier, and game information to generate a target corpus.

[0021] Specifically, acquiring text corpora from multiple game communities and generating a target corpus is a fundamental processing step for subsequent tasks such as adjacent character cohesion calculation, low-cohesion segmentation location identification, and cross-community common word filtering. Because text sources in game communities are scattered, colloquial, and contain numerous aliases and abbreviations, directly analyzing the original text can easily lead to the same entity being split into multiple unrelated expressions, or allow noise such as emoticons, links, and spam phrases to enter the subsequent cohesion calculation process, thus affecting the generation of candidate stable segments and candidate word sets. Therefore, this embodiment first uniformly collects, standardizes, and hierarchically organizes the multi-source community text, and then outputs a target corpus that can be directly used in subsequent mining processes.

[0022] In practice, the system first collects post titles, post texts, replies, strategy comments, bullet screen texts, battle report descriptions, and event discussion texts from multiple game communities. For each text, it simultaneously extracts the posting time, community source, section identifier, interaction popularity marker, and game information. The community source distinguishes between different sources such as forums, video comment communities, and live streaming communities. The section identifier distinguishes between character discussion areas, equipment discussion areas, version exchange areas, and event areas. The interaction popularity marker can be generated from at least one of the following: number of replies, likes, favorites, or shares. This method ensures that each text carries clear source and game attributes before entering subsequent processing, providing a basis for subsequent statistical analysis of the distribution of candidate keywords across different communities and games.

[0023] After the initial data collection, the text undergoes unified cleaning and standardization. Specifically, completely or highly repetitive forwarded text and copied comments are merged; links, emoticons, meaningless emoticons, and continuous spam symbols are removed; continuously repeating characters are compressed into a standard form within a preset length; and full-width and half-width characters, simplified and traditional Chinese characters, and uppercase and lowercase forms are uniformly converted. Simultaneously, a preset deactivated noise fragment library is used to remove noise expressions that do not carry entity recognition value, such as "like," "just passing by," "asking for help," and "rushing forward," resulting in standardized text fragments. For mixed symbol strings containing version numbers, season numbers, region codes, equipment numbers, and character abbreviations that have identification value in game scenarios, they are not removed as noise but instead enter a targeted retention process to avoid mistakenly deleting key expressions that distinguish the community.

[0024] Furthermore, special expressions within standardized text fragments undergo normalization processing. Preset retention rules preserve expressions containing combinations of numbers and text, such as version 3.2, season 12, region 1, and equipment number 1023. Alias ​​unification rules map the official name, abbreviation, and colloquialisms of the same game entity to a unified identifier. For example, in an open-world game community, players might use the official character name in post titles, a two-character abbreviation in replies, and a nickname in bullet comments to refer to the same character. The system can unify these multiple expressions under the same candidate entity name based on a preset alias dictionary, bracket interpretation patterns, and co-occurrence relationships within the same post. Subsequently, a hierarchical association index is established by combining community origin, section identifier, and game information. A multi-level index structure—game—community—section—text—is optimally constructed, enabling subsequent processing to compare term distribution across different communities within the same game and identify common terms across different games.

[0025] For example, in a specific instance, the system simultaneously collects text corpora from Game A and Game B communities. In Game A, the version discussion section contains numerous posts about exploring the new map in version 3.2, equipment builds for fire-type main characters, and event dungeons in server 1. In Game B, the season discussion section contains numerous posts about team compositions for climbing the ranks in season 12, the nerf to item #1023, and strategies for exploring new servers. The system first extracts the posting time, source site, discussion forum, and game of each text. Then, it cleans up links, emoticons, and duplicate content, unifying "version 3.2" as "version 3.2," unifying traditional Chinese characters as simplified Chinese, and grouping the official name, abbreviation, and nickname of the same character under a unified entity identifier.

[0026] Next, a hierarchical association index is established based on the game, community origin, and section identifier, and the processed text fragments are encapsulated into a target corpus. Subsequently, when performing cohesion statistics, the connection relationship between adjacent characters can be accurately identified based on this target corpus, and when performing cross-community public word filtering, it can distinguish between public expressions such as "strategy guide" and "ranking up" and exclusive entity expressions that are only concentrated in specific game communities.

[0027] Through the above implementation methods, multi-source game community texts can be uniformly collected, cleaned, standardized, and hierarchically organized, enabling the unified merging of multiple expressions of the same entity. This provides a structurally consistent and clearly cited target corpus foundation for subsequent low-cohesion reverse segmentation and cross-community public word identification, thereby improving the stability and processing accuracy of the subsequent entity word mining process.

[0028] In some embodiments, based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus, the cohesion corresponding to the character connection position is calculated, including: For each text segment in the target corpus, perform adjacent scanning according to character order, extract each adjacent character pair, and count the number of adjacent co-occurrences of each adjacent character pair and the number of occurrences of each single character to establish character adjacency statistics results; Based on the character adjacency statistics and the total number of characters in the target corpus, the joint probability of each adjacent character pair and the marginal probability of each single character are determined. Low-frequency adjacent character pairs are filtered out by frequency lower limit or smoothed to generate an effective probability distribution. According to the preset cohesion calculation rules, the ratio of the joint probability of each adjacent character pair to the marginal probability of the corresponding single character is calculated to determine the cohesion corresponding to the character connection position of each adjacent character pair, and the cohesion is written into the character connection position marker sequence of the corresponding text segment.

[0029] Specifically, calculating the cohesion of character connection positions based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus is a step in the character-level connection strength modeling process of the target corpus, following the aforementioned completion of multi-game community text collection, noise reduction and normalization, hybrid symbol normalization, and hierarchical association index construction. This step does not directly determine whether a string is an entity word, but rather establishes a quantitative connection strength for each character connection position within a text segment. This provides foundational data for subsequently determining a preset cohesion threshold based on the overall corpus cohesion distribution, identifying low-cohesion segmentation positions, and performing reverse segmentation.

[0030] In the specific implementation, the standardized text segments in the target corpus are first scanned sequentially according to character order. For each text segment, all adjacent character pairs are extracted based on their preceding and following adjacency relationships. The co-occurrence count of each adjacent character pair and the occurrence count of each individual character in the target corpus are then accumulated to form character adjacency statistics. The co-occurrence count characterizes the strength of two characters appearing consecutively in actual community expressions, while the occurrence count of a single character characterizes its own marginal activity level. Furthermore, to ensure consistency in statistical criteria across different community sources and game identifiers, a unified traversal is performed on each text segment based on the aforementioned game-community-section-text hierarchical association index. The resulting character adjacency statistics are then written into the corresponding statistical cache structure to support subsequent joint probability and marginal probability calculations across the entire corpus.

[0031] After obtaining the character adjacency statistics, the joint probability of each adjacent character pair and the marginal probability of each individual character are determined by combining the total number of characters in the target corpus. For adjacent character pairs that occur too infrequently, it is preferable to perform frequency lower limit screening to avoid the extremely low-frequency accidental splicing causing disturbance to the cohesion calculation results; for low-frequency character pairs that need to be retained but have potential recognition value, a smoothing correction method can be used to generate an effective probability distribution.

[0032] For example, smoothing correction can be achieved by introducing a preset compensation amount for the number of adjacent co-occurrences and the number of single character occurrences, so that expressions such as version numbers, season numbers, region codes, and character abbreviations, which have specific semantics in the game community but are not frequently used in the early stages, are not directly excluded. Then, according to the preset cohesion calculation rules, the joint probability of each pair of adjacent characters is calculated as a ratio to the marginal probability of the corresponding single character to obtain the cohesion corresponding to each character connection position, and the cohesion is written into the character connection position marker sequence of the corresponding text segment in character order.

[0033] For example, in a specific example, targeting a corpus comprised of discussion forums for game A (version A), character guides, and game B (season B), the system first performs adjacent scans on standardized text fragments such as "Fire Main character early game in version 3.2," "Fire Main character equipment recommendations," "Season 12 early game lineup," and "Equipment nerf for item #1023." In the "Fire Main character equipment recommendations," adjacent character pairs such as "Fire Main," "Main character," "equipment," "equipment recommendation," and "recommendation" are extracted, and the frequency of single characters such as "Fire," "Main character," "equipment," and "equipment" is counted separately. In the "Season 12 early game lineup," mixed expressions related to the season number, such as "12," "2 season," and "season," are retained and included in the unified statistics.

[0034] Subsequently, the system calculates the joint probability of each adjacent character pair and the marginal probability of each individual character based on the total number of characters in the target corpus. If "Fire" and "Master" appear consecutively in multiple communities of Game A, and their individual distributions are not uniform, the cohesion of the corresponding character connection position is high. If "Package" appears stably in multiple texts, the cohesion between "Package" and "Package" is also high. However, for connections formed by "Recommendation" and other characters before and after, if they only appear occasionally in a small number of texts, their cohesion is relatively low. The system writes the above calculation results into a character connection position marker sequence according to the original text character order. Subsequently, it can identify low-cohesion segmentation positions based on the marker sequence and a preset cohesion threshold, and then perform reverse segmentation and candidate stable segment generation based on this.

[0035] Through the above implementation methods, based on the aforementioned target corpus construction results, a unified and quantifiable cohesion representation can be established for the connection positions of each character within the game community text. This not only preserves the connection information unique to game scenarios, such as version numbers, season numbers, and character abbreviations, but also suppresses the interference of low-frequency accidental splicing on subsequent segmentation processing, thereby providing a stable data foundation for the identification of low-cohesion segmentation positions and the extraction of candidate stable segments.

[0036] In some embodiments, a preset cohesion threshold is determined based on the cohesion distribution of the entire corpus, and low-cohesion segmentation locations are identified based on the preset cohesion threshold, including: The cohesion of each character connection position in the target corpus is aggregated. A cohesion statistical sequence of the entire corpus is established according to community and game identifiers. The boundaries of low quantile intervals, distribution density inflection points and significance boundaries are extracted to generate an initial threshold candidate set. Based on preset task constraints, interval filtering and parameter verification are performed on the initial threshold candidate set. The preset task constraints include at least one of the following: fragment compression rate after reverse segmentation, candidate stable fragment length distribution, candidate word retention ratio, and community distribution consistency. A globally applicable preset cohesion threshold is determined. The cohesion of each character connection position is compared with a preset cohesion threshold, and the character connection positions with a cohesion lower than the preset cohesion threshold are marked as low cohesion split positions.

[0037] Specifically, in this embodiment, determining a preset cohesion threshold based on the cohesion distribution of the entire corpus, and identifying low-cohesion segmentation positions based on the preset cohesion threshold, is a process of globally calibrating the strength of character connections, building upon the previously completed construction of the target corpus and the calculation of cohesion at each character connection position. The core of this step is not to temporarily set a segmentation standard for a single text, but rather to uniformly aggregate the cohesion results from multiple game communities, multiple sections, and multiple game identifiers to construct a globally universal threshold that can be reused across communities, ensuring consistency in the subsequent reverse segmentation process across different game corpora. Since game communities contain character abbreviations, equipment numbers, event names, and colloquial expressions, setting the threshold too low may retain a large number of invalid connections; setting the threshold too high may incorrectly segment stable expressions with entity recognition value, such as "Fire Master," "Version 3.2," and "Season 12." Therefore, this embodiment uses a combination of cohesion distribution statistics and task constraint verification to determine the preset cohesion threshold.

[0038] In practical implementation, the cohesion of all character connection positions in the target corpus is first aggregated, and a corpus-wide cohesion statistical sequence is established according to community and game identifiers. Specifically, the character connection position cohesion obtained from game A's version discussion forum, character strategy forum, event forum, and game B's season discussion forum, equipment discussion forum, and battle report comment forum are uniformly written into the statistical pool, and then local distribution subsequences and global distribution main sequences are generated according to game and community identifiers, respectively. Based on this, distribution analysis is performed on the corpus-wide cohesion statistical sequence to extract low quantile interval boundaries, distribution density inflection points, and significance boundaries, forming an initial threshold candidate set. Among them, the low quantile interval boundaries are used to give the approximate range of weak connection regions, the distribution density inflection points are used to identify the transition boundary from high-frequency weak connections to stable connections, and the significance boundaries are used to identify statistical boundaries that deviate from the random splicing state. The above multiple candidate values ​​together constitute the threshold search interval, rather than directly using a single empirical value, thus making the threshold generation process data-driven.

[0039] After obtaining the initial candidate threshold set, further interval filtering and parameter verification are performed based on preset task constraints. The preset task constraints include at least one of the following: fragment compression rate after reverse segmentation, distribution of candidate stable fragment lengths, proportion of candidate words retained, and consistency of community distribution. Specifically, if a candidate threshold causes a large amount of text to be segmented into too many single-character fragments, it indicates that the threshold is too high; if a candidate threshold causes an abnormally high proportion of long fragments and an excessively high proportion of candidate words retained after segmentation, it indicates that the threshold is too low.

[0040] Furthermore, the aforementioned standardized game-specific expressions can be used for verification. For example, segments such as "fire master," "equipment," "season," and "exploration" that appear consistently in multiple texts should not be mis-segmented on a large scale, while boundaries of weakly connected elements such as "recommend," "how," and "what" are more easily identified as candidate segmentation positions. Based on the above constraints, after filtering the initial threshold candidate set, a globally universal preset cohesion threshold that can be reused in multiple game community scenarios is determined.

[0041] After determining the threshold, the cohesion of each character connection position is compared with a preset cohesion threshold, and character connection positions with a cohesion lower than the threshold are marked as low-cohesion segmentation positions. For example, in the text "Version 3.2 Fire Main's Early Game Development" in the discussion forum of Game A, after the aforementioned cohesion calculation, the cohesion of internal connection positions such as 3.2, version, Fire Main, and Early Game Development is usually high, while the cohesion of connection positions located at the boundaries of different semantic units is relatively low. The system can then mark positions with a cohesion lower than the preset threshold as segmentation positions. In the text "Season 12 Early Game Lineup and Equipment 1023 Nerf" in the discussion forum of Game B, the system also uses a unified threshold to identify weak connection boundaries and writes the identification results into the character connection position marking sequence of the corresponding text segment. When performing reverse segmentation later, the text can be directly broken at low-cohesion positions based on this marking sequence to generate a set of candidate stable segments without having to repeatedly estimate the segmentation criteria for each text.

[0042] Through the above implementation methods, a globally universal preset cohesion threshold with statistical basis and task constraint support can be formed based on the cohesion distribution of the entire corpus. Based on this, low cohesion segmentation positions in different game community texts can be stably identified, thereby improving the consistency of the reverse segmentation standard, reducing the fluctuations caused by empirical thresholding, and providing a reliable foundation for subsequent candidate stable segment extraction and entity word mining.

[0043] In some embodiments, the target corpus is reverse-segmented based on low-cohesion segmentation positions to obtain a set of candidate stable segments, and a set of candidate words is assembled based on the adjacency relationships and length constraints of the candidate stable segment set, including: The text segments are segmented at low cohesion locations to generate corresponding segment sequences. Segments with a length of 1 and a frequency below a preset frequency threshold are removed, retained, or merged to obtain an initial stable segment sequence. Identify mixed symbol fragments containing version number, season number, region code, equipment number or character abbreviation in the initial stable fragment sequence, and perform targeted retention by combining preset retention rules, alias normalization rules and adjacent context constraints to generate a candidate stable fragment set; Based on the sequential adjacency relationship of the candidate stable fragment set in the original text, the cross-text repetition and co-occurrence relationship, and the preset length constraint, a restricted assembly and candidate screening are performed to obtain the candidate word set.

[0044] Specifically, this embodiment first performs breakpoint segmentation on each text segment according to the low cohesion segmentation positions, generating a corresponding segment sequence. Specifically, the character connection position marker sequence in each text segment is compared with a preset cohesion threshold. Segmentation breakpoints are inserted at positions below the threshold, thereby converting the original text into multiple sequentially arranged segment units. For segments with a length of 1 after segmentation and an occurrence frequency below a preset frequency threshold, the system does not directly retain them all. Instead, it performs differentiated processing based on their statistical characteristics and contextual state: single-character segments that clearly lack recognition value are directly eliminated; segments that are single characters but located in high-frequency game contexts are given a delayed retention mark; and single-character segments that may form entity expressions together with preceding and following segments are given a mergeable mark. This yields an initial stable segment sequence, providing a foundation for subsequent special segment identification and restricted assembly.

[0045] After generating the initial stable fragment sequence, further identification is performed on mixed symbol fragments containing version numbers, season numbers, region codes, equipment numbers, or character abbreviations. These fragments are then targeted for retention based on preset retention rules, alias unification rules, and adjacent context constraints. Preset retention rules prevent fragments from being mistakenly deleted due to containing numbers or symbols, such as those from version 3.2, season 12, region 1, or equipment number 1023. Alias ​​unification rules map official names, two-letter abbreviations, and community nicknames to the same candidate entity channel. Adjacent context constraints determine whether a fragment consistently appears in game contexts such as early game progression, gear setups, nerfs, event dungeons, and seasonal team compositions. After these processes, a set of candidate stable fragments is obtained. It should be noted that the candidate stable fragments obtained here are still candidate units, not final entity words. Their purpose is to provide clear-bounded and semantically stable intermediate representations for subsequent constrained assembly.

[0046] After the candidate stable fragment set is formed, the system performs restricted assembly and candidate filtering based on the sequential adjacency relationship of each fragment in the original text, cross-text repetition and co-occurrence relationship, and preset length constraints to obtain a candidate word set. Sequential adjacency relationship is used to determine whether two or more candidate stable fragments appear consecutively in the text; cross-text repetition and co-occurrence relationship is used to determine whether the same assembly result appears repeatedly in multiple post titles, body texts, replies, or bullet comments; preset length constraints are used to limit the maximum and minimum length of candidate words to avoid generating excessively long mixed fragments or excessively short invalid fragments. Assembly results that do not meet the repetition and co-occurrence conditions, length constraints, or contextual consistency conditions are eliminated or downgraded by the system; assembly results that meet the conditions are added to the candidate word set.

[0047] For example, in a specific instance, the game version discussion forum and character strategy forum contain texts such as "3.2 version Fire Main early game guide," "Fire Main equipment recommendations," and "Fire Main dungeon strategies." The game season discussion forum and equipment discussion forum contain texts such as "Season 12 early game team composition," "Equipment 1023 nerf," and "Region 1 event dungeon." After the system performs breakpoint segmentation on the "3.2 version Fire Main early game guide" based on the aforementioned low-cohesion segmentation locations, it obtains initial fragments such as "3.2 version," "Fire Main," and "early game guide." For the "Season 12 early game team composition," it obtains initial fragments such as "Season 12," "early game guide," and "team composition." For the "Equipment 1023 nerf," it obtains initial fragments such as "Equipment 1023" and "nerf." For the mixed symbol fragments such as "3.2 version," "Season 12," and "Equipment 1023," the system directly retains them according to preset retention rules. For the "Fire Main" character abbreviation fragment, it retains it based on the aforementioned alias unification results and the high-frequency co-occurrence relationships in the character strategy forum.

[0048] Subsequently, the system performs restricted assembly based on the stable adjacency relationships between the Fire Main and the terms "exploration," "gear setup," and "dungeon" across multiple texts, generating candidate terms such as "Fire Main exploration" and "Fire Main gear setup." Based on the recurring co-occurrence relationship between "Season 12" and "exploration lineups" in the season discussion forum, candidate expressions for "Season 12 exploration lineups" are generated. After length constraints and recurring co-occurrence filtering, the retained assembly results are written into a candidate word set for subsequent cross-community public word filtering and entity confirmation processing.

[0049] Through the above implementation method, invalid connections with weak internal cohesion can be eliminated first based on the low cohesion segmentation position. Then, mixed symbol fragments and abbreviation fragments with recognition value in the game scene are selectively retained. Candidate words are assembled in a restricted manner by using sequential adjacency relationship, cross-text repetition co-occurrence relationship and length constraint, thereby reducing redundant combinations in the candidate generation process, improving the stability and effectiveness of the candidate word set, and providing a reliable foundation for subsequent cross-community public word recognition and target entity word mining.

[0050] In some embodiments, cross-community public words are identified based on community coverage, game coverage, and distribution dispersion, and cross-community public words are filtered out from the candidate word set, including: Statistically analyze the frequency of each candidate word in the candidate word set under different community identifiers and different game identifiers, and construct the community distribution matrix and game distribution matrix corresponding to the candidate words; Based on the community distribution matrix and the game distribution matrix, the community coverage, game coverage, frequency ratio difference and distribution entropy value corresponding to each candidate word are determined, and a cross-community distribution representation is formed. The cross-community distribution representation is matched with the preset public word judgment conditions. Candidate words that meet the conditions of game coverage greater than the preset coverage threshold and distribution entropy value greater than the preset discrete threshold are marked as cross-community public words and filtered out from the candidate word set.

[0051] Specifically, in this embodiment, identifying cross-community common words based on community coverage, game coverage, and distribution dispersion, and filtering out cross-community common words from the candidate word set, is a process of cross-community discriminative judgment on candidate words after the aforementioned completion of target corpus construction, character connection position cohesion calculation, low-cohesion segmentation position identification, and candidate word set generation. The purpose of this step is not to re-determine whether candidate words possess internal connectivity stability, but rather to further identify generalized expressions that, while satisfying cohesion and frequency conditions in local text, commonly appear in multiple communities and games and lack community specificity. This embodiment filters out cross-community common words by constructing distribution representations of candidate words under different community and game identifiers.

[0052] In practical implementation, the frequency of each candidate word in the candidate word set under different community and game identifiers is first counted, and a community distribution matrix and a game distribution matrix corresponding to the candidate words are constructed. The community distribution matrix records the distribution of the same candidate word across different community sources such as character discussion areas, equipment discussion areas, version discussion areas, event areas, battle report comment areas, and bullet screen interaction areas; the game distribution matrix records the frequency distribution of the same candidate word in game A, game B, and other game corpora. Since the aforementioned target corpus has already established a hierarchical association index of game-community-section-text, this step can directly trace back to the game and community to which the candidate word belongs by its identifier, and accumulate cross-community statistical results for subsequent analysis.

[0053] After obtaining the community distribution matrix and game distribution matrix, the community coverage, game coverage, frequency ratio difference, and distribution entropy value corresponding to each candidate word are further determined, forming a cross-community distribution representation. Specifically, the community coverage represents the number of community sources or section types in which the candidate word appears; the game coverage represents the number of different games in which the candidate word appears; the frequency ratio difference measures the concentration difference of the candidate word between the primary and secondary distribution games; and the distribution entropy value represents the evenness of the candidate word's distribution across multiple games or communities.

[0054] For example, if a candidate word appears frequently in multiple games and the frequency difference between games is small, its distribution entropy value is usually high and the frequency ratio difference is usually low, which can be regarded as lacking game differentiation ability; if a candidate word is mainly concentrated in a single game or a small number of related communities, its game coverage is small or the frequency ratio difference is large, which can be regarded as having strong community specificity.

[0055] Based on this, the cross-community distribution representation is matched with preset public word determination criteria. These criteria include at least a game coverage area greater than a preset coverage threshold and a distribution entropy value greater than a preset discrete threshold. If necessary, further verification can be performed by combining this with a frequency percentage difference value less than a preset difference threshold. Candidate words that meet the criteria are marked as cross-community public words and filtered out from the candidate word set. Candidate words that do not meet the criteria but exhibit significant concentrated distribution characteristics in a specific game or community are retained for subsequent entity confirmation and ranking based on game context trigger features and contextual co-occurrence relationships.

[0056] For example, in a specific instance, the system performs joint statistics on the candidate word sets generated from the version discussion forum and character strategy forum of game A, and the season discussion forum and equipment discussion forum of game B. The statistical results show that candidate words such as "strategy," "ranking up," "early game exploration," and "how to play" appear simultaneously in multiple communities of both game A and game B, and their frequency distribution across different games is relatively balanced. Therefore, these words have a large game coverage and a high distribution entropy value, and the system marks them as cross-community common words and filters them out.

[0057] In contrast, candidate terms such as "Fire Master," "Version 3.2 map," "Season 12 lineup," and "Equipment #1023" primarily appear in the character, version, or season sections of their respective games, with significantly lower frequency of appearance in other games. Therefore, their frequency distribution differs greatly, or their game coverage is limited, leading the system to retain them in the candidate term set. This method allows the system to distinguish between generalized expressions that satisfy connection stability but lack community specificity and candidate entity expressions that truly possess game-distinguishing value.

[0058] Through the above implementation methods, a cross-community distribution representation can be formed based on the distribution of candidate words under different community identifiers and different game identifiers. Based on this representation, cross-community common words can be identified and filtered out, thereby reducing the interference of generalized expressions on the candidate word set, improving the community specificity and entity relevance of retained candidate words, and providing a more accurate input basis for subsequent entity confirmation and ranking.

[0059] In some embodiments, the filtered candidate words are combined with game context trigger features and context co-occurrence relationships to perform entity confirmation and ranking, and output the entity word mining results corresponding to the target game community, including: Extract the context window of each filtered candidate word in the target corpus, identify the role, equipment, dungeon, event, version, season and gameplay trigger words associated with the candidate words, and generate the game context trigger representation corresponding to the candidate words; Based on the game context triggering representation and the co-occurrence objects, co-occurrence frequency and positional adjacency relationships in the context window, a context co-occurrence representation corresponding to the candidate words is constructed, and the entity confidence corresponding to each candidate word is determined accordingly. The filtered candidate words are sorted and selected according to entity confidence and preset sorting rules to generate entity word mining results corresponding to the target game community.

[0060] Specifically, this embodiment no longer determines whether candidate words should be retained based on the strength of internal character connections or cross-community distribution. Instead, it further determines whether the candidate word is in a stable game entity context and whether a continuous and reusable co-occurrence relationship exists between the candidate word and game objects such as characters, equipment, dungeons, events, versions, seasons, and gameplay. To this end, this embodiment extracts a context window for each candidate word, identifies trigger words that can characterize game entity attributes within the context window, and constructs a context co-occurrence representation by combining co-occurrence objects, co-occurrence frequency, and positional adjacency relationships. This generates entity confidence and outputs the entity word mining results corresponding to the target game community.

[0061] In the specific implementation, the context window is first extracted from the target corpus around each filtered candidate word. The context window can be the text content within a preset character range, preset word range, or preset sentence range before and after the candidate word. Then, the context window is used to identify the character trigger words, equipment trigger words, dungeon trigger words, event trigger words, version trigger words, season trigger words, and gameplay trigger words associated with the candidate words, so as to generate the game context trigger representation corresponding to the candidate words.

[0062] Among them, character trigger words can include team composition, talents, constellations, and character level; equipment trigger words can include unique weapons, stats, enhancement, and forging; dungeon trigger words can include drops, clearing, bosses, and challenges; event trigger words can include reruns, limited-time events, exchanges, and rewards; version trigger words can include updates, nerfs, buffs, and reworks; season trigger words can include seasons, ranks, ranking up, and results; and gameplay trigger words can include early game exploration, team compositions, and mechanics. Through the identification of these trigger words, different candidate words can be mapped to different game entity context channels, thus providing semantic basis for subsequent entity confirmation.

[0063] After obtaining the game context trigger representation, a context co-occurrence representation is further constructed by combining the co-occurrence objects, co-occurrence frequencies, and positional adjacency relationships in the context window. Specifically, co-occurrence objects are used to record which identified objects or high-confidence terms frequently co-occur with candidate words; co-occurrence frequencies are used to count the degree of repetition of this co-occurrence relationship in the target corpus; and positional adjacency relationships are used to determine whether candidate words, trigger words, and co-occurrence objects appear in a relatively stable order and distance.

[0064] For example, if a candidate word frequently co-occurs with trigger words like "equipment dungeon" or "exploration" in adjacent sentence segments, it can be determined that it is more likely to belong to a character or gameplay entity; if a candidate word frequently appears adjacent to expressions like "word entry," "enhancement," or "number 1023," it can be determined that it is more likely to belong to an equipment entity. Based on game context trigger representations and contextual co-occurrence representations, the system can generate entity confidence for each candidate word. Entity confidence can comprehensively consider at least two of the following: the degree of trigger word matching, the stability of co-occurring objects, the degree of cross-textual repetition and co-occurrence, and the stability of positional adjacency, to avoid identifying entities based on a single clue.

[0065] For example, in a specific instance, the system performs entity verification on candidate words retained in the game A version discussion forum, character strategy forum, and game B season discussion forum and equipment discussion forum. For the candidate word "Fire Main" in game A, the system identifies trigger words such as equipment, dungeon, exploration, and version in context windows such as "Fire Main equipment recommendation," "Fire Main dungeon strategy," and "Fire Main early game." It also finds that "Fire Main" co-occurs with objects such as lineup, level, and team composition, and in most cases, it is located before the trigger words or in adjacent phrases. Therefore, a high entity confidence score is assigned to "Fire Main."

[0066] For the candidate term "Version 3.2 map," the system identified trigger words such as "version," "update," "map," and "mechanics" in texts related to the new map exploration and map mechanism adjustments after the version update in Version 3.2. It also confirmed that these trigger words primarily appeared in the Game A version discussion forum, thus identifying them as version-related entities. For the candidate term "Equipment No. 1023" in Game B, the system identified trigger words such as "equipment," "nerf," "attribute," and "enhancement" in texts such as "Equipment No. 1023 nerf" and "Equipment No. 1023 affix recommendation." Furthermore, this candidate term consistently co-occurred with elements like "seasonal lineup" and "ranking strategies," thus assigning it a high confidence level as an equipment-related entity. However, for some candidate terms that, while not identified as cross-community public terms, lacked object trigger relationships due to emotional expressions like "truly outrageous" and "too difficult" in the context, a lower entity confidence level was assigned, and they were downgraded during the sorting phase. Finally, the system sorts and filters the retained candidate words according to entity confidence and preset sorting rules, and outputs entity word mining results such as Fire Master and version 3.2 map in Game A community and Season 12 lineup and equipment No. 1023 in Game B community.

[0067] Through the above implementation method, after the candidate word filtering is completed, the candidate words can be further combined with game context trigger features and context co-occurrence relationships to perform entity confirmation and sorting, so that candidate words that are truly in a stable game entity context are given priority retention, thereby reducing the interference of residual generalized expressions on the result output and improving the accuracy and usability of the target game community entity word mining results.

[0068] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0069] Figure 2 This is a schematic diagram of the structure of the game community entity word mining device provided in this application embodiment. For example... Figure 2 As shown, the entity word mining device for the game community includes: The acquisition module 201 is used to acquire text corpora corresponding to multiple game communities, perform noise reduction and normalization, mixed symbol normalization, and association processing of community identifiers and game identifiers on the text corpora, and generate a target corpus set. The calculation module 202 is used to calculate the cohesion corresponding to the character connection position based on the co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus, and to determine the preset cohesion threshold according to the cohesion distribution of the whole corpus, and to identify low cohesion segmentation positions based on the preset cohesion threshold. The segmentation module 203 is used to perform reverse segmentation on the target corpus based on the low cohesion segmentation position to obtain a set of candidate stable segments, and assemble a set of candidate words based on the adjacency relationship and length constraints of the set of candidate stable segments. The identification module 204 is used to count the frequency distribution of each candidate word in multiple game communities and multiple game identifiers, identify cross-community public words based on community coverage, game coverage and distribution dispersion, and filter out cross-community public words from the candidate word set; Output module 205 is used to perform entity confirmation and ranking on the filtered candidate words by combining game context trigger features and context co-occurrence relationships, and output the entity word mining results corresponding to the target game community.

[0070] In some embodiments, Figure 2 The acquisition module 201 collects post titles, texts, replies, strategy comments, bullet screen texts, battle report descriptions, and activity discussion texts from various game communities, extracting the posting time, community source, section identifier, interaction popularity marker, and game information corresponding to each text. It then performs duplicate content merging, link and emoticon cleaning, abnormal character compression, full-width / half-width character unification, simplified / traditional character conversion, and deactivation of noisy segments to obtain standardized text fragments. Finally, it identifies mixed symbol strings such as version numbers, season numbers, region codes, equipment numbers, and character abbreviations in the standardized text fragments, standardizes them according to preset retention rules and alias unification rules, and establishes a hierarchical association index based on community source, section identifier, and game information to generate the target corpus.

[0071] In some embodiments, Figure 2 The calculation module 202 performs adjacent scanning on each text segment in the target corpus according to the character order, extracts each adjacent character pair, and counts the number of adjacent co-occurrences of each adjacent character pair and the number of occurrences of each individual character, establishing character adjacency statistics. Based on the character adjacency statistics and the total number of characters in the target corpus, it determines the joint probability corresponding to each adjacent character pair and the marginal probability corresponding to each individual character, and performs frequency lower limit screening or smoothing correction on low-frequency adjacent character pairs to generate an effective probability distribution. According to the preset cohesion calculation rules, it performs ratio calculation on the joint probability of each adjacent character pair and the marginal probability of the corresponding individual character to determine the cohesion corresponding to the character connection position of each adjacent character pair, and writes the cohesion into the character connection position marker sequence of the corresponding text segment.

[0072] In some embodiments, Figure 2The calculation module 202 aggregates the cohesion corresponding to each character connection position in the target corpus, establishes a cohesion statistical sequence of the entire corpus according to community identifiers and game identifiers, and extracts the low quantile interval boundaries, distribution density inflection points, and saliency boundaries to generate an initial threshold candidate set. Based on preset task constraints, interval filtering and parameter verification are performed on the initial threshold candidate set. The preset task constraints include at least one of the following: fragment compression rate after reverse segmentation, candidate stable fragment length distribution, candidate word retention ratio, and community distribution consistency. A globally applicable preset cohesion threshold is determined. The cohesion corresponding to each character connection position is compared with the preset cohesion threshold, and character connection positions with less than the preset cohesion threshold are marked as low cohesion segmentation positions.

[0073] In some embodiments, Figure 2 The segmentation module 203 performs breakpoint segmentation on each text segment according to the low cohesion segmentation position, generates the corresponding segment sequence, and performs elimination, retention marking or merging marking on segments with a length of 1 after segmentation and an occurrence frequency lower than the preset frequency threshold to obtain an initial stable segment sequence; identifies mixed symbol segments containing version number, season number, region code, equipment number or character abbreviation in the initial stable segment sequence, and performs targeted retention in combination with preset retention rules, alias normalization rules and adjacent context constraints to generate a candidate stable segment set; performs restricted assembly and candidate screening based on the sequential adjacency relationship, cross-text repetition co-occurrence relationship and preset length constraints of the candidate stable segment set in the original text to obtain a candidate word set.

[0074] In some embodiments, Figure 2 The identification module 204 counts the frequency of each candidate word in the candidate word set under different community identifiers and different game identifiers, and constructs the community distribution matrix and game distribution matrix corresponding to the candidate words. Based on the community distribution matrix and game distribution matrix, it determines the community coverage, game coverage, frequency ratio difference and distribution entropy value corresponding to each candidate word, and forms a cross-community distribution representation. The cross-community distribution representation is matched with the preset public word judgment conditions. Candidate words that meet the conditions of game coverage greater than the preset coverage threshold and distribution entropy value greater than the preset discrete threshold are marked as cross-community public words and filtered out from the candidate word set.

[0075] In some embodiments, Figure 2The output module 205 extracts the context window of each filtered candidate word in the target corpus, identifies the role, equipment, dungeon, event, version, season and gameplay trigger words associated with the candidate words, and generates the game context trigger representation corresponding to the candidate words; based on the game context trigger representation and the co-occurrence objects, co-occurrence frequency and positional adjacency relationship in the context window, it constructs the context co-occurrence representation corresponding to the candidate words, and determines the entity confidence corresponding to each candidate word accordingly; the filtered candidate words are sorted and filtered according to the entity confidence and preset sorting rules to generate entity word mining results corresponding to the target game community.

[0076] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0077] Figure 3 This is a schematic diagram of the structure of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various device embodiments described above.

[0078] For example, computer program 303 may be divided into one or more modules / units, which are stored in memory 302 and executed by processor 301 to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 303 in electronic device 3.

[0079] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0080] Processor 301 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0081] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or RAM. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 302 can include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0082] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0083] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0084] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] In the embodiments provided in this application, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of apparatus or units may be electrical, mechanical, or other forms.

[0086] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0087] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0089] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although the technical solutions of this application are described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for mining entity words in a game community, characterized in that, include: Obtain text corpora corresponding to multiple game communities, and perform noise reduction and normalization, mixed symbol normalization, and association processing of community identifiers and game identifiers on the text corpora to generate a target corpus set; Based on the co-occurrence statistics and edge probability distribution of adjacent characters in the target corpus, the cohesion corresponding to the character connection position is calculated, and a preset cohesion threshold is determined according to the cohesion distribution of the entire corpus. Low cohesion segmentation positions are identified based on the preset cohesion threshold. The target corpus is reverse-segmented based on the low-cohesion segmentation position to obtain a set of candidate stable segments, and a set of candidate words is assembled based on the adjacency relationship and length constraints of the set of candidate stable segments. The frequency distribution of each candidate word in multiple game communities and multiple game identifiers is statistically analyzed. Cross-community common words are identified based on community coverage, game coverage, and distribution dispersion. The cross-community common words are then filtered out from the candidate word set. After filtering out candidate words, entity confirmation and ranking are performed by combining game context trigger features and context co-occurrence relationships, and the entity word mining results corresponding to the target game community are output.

2. The method according to claim 1, characterized in that, The process involves acquiring text corpora corresponding to multiple game communities, performing noise reduction and normalization, mixed symbol normalization, and association processing between community identifiers and game identifiers on the text corpora to generate a target corpus set, including: Collect post titles, texts, replies, strategy comments, bullet screen texts, battle report descriptions, and event discussion texts from various game communities, and extract the posting time, community source, section identifier, interaction popularity marker, and game information corresponding to each text. The text is processed by merging duplicate content, cleaning links and emojis, compressing abnormal characters, unifying full-width and half-width characters, converting between traditional and simplified Chinese characters, and removing noisy segments to obtain standardized text segments. The system identifies mixed symbol strings containing version numbers, season numbers, region codes, equipment numbers, and character abbreviations in the standardized text fragments, normalizes them according to preset retention rules and alias unification rules, and establishes a hierarchical association index based on the community source, section identifier, and game information to generate the target corpus.

3. The method according to claim 1, characterized in that, The calculation of co-occurrence statistics and marginal probability distribution of adjacent characters in the target corpus, including: For each text segment in the target corpus, perform adjacent scanning according to character order, extract each adjacent character pair, and count the number of adjacent co-occurrences of each adjacent character pair and the number of occurrences of each single character to establish character adjacency statistics results; Based on the character adjacency statistics and the total number of characters in the target corpus, the joint probability of each adjacent character pair and the marginal probability of each single character are determined, and low-frequency adjacent character pairs are subjected to frequency lower limit screening or smoothing correction to generate an effective probability distribution. According to the preset cohesion calculation rules, the ratio of the joint probability of each adjacent character pair to the marginal probability of the corresponding single character is calculated to determine the cohesion corresponding to the character connection position of each adjacent character pair, and the cohesion is written into the character connection position marker sequence of the corresponding text segment.

4. The method according to claim 1, characterized in that, The step of determining a preset cohesion threshold based on the cohesion distribution of the entire corpus, and identifying low-cohesion segmentation locations based on the preset cohesion threshold, includes: The cohesion of each character connection position in the target corpus is aggregated, and a cohesion statistical sequence of the entire corpus is established according to the community identifier and game identifier. The low quantile interval boundary, distribution density inflection point position and saliency boundary position are extracted to generate an initial threshold candidate set. Based on preset task constraints, interval filtering and parameter verification are performed on the initial threshold candidate set. The preset task constraints include at least one of the following: fragment compression rate after reverse segmentation, candidate stable fragment length distribution, candidate word retention ratio, and community distribution consistency, to determine a globally applicable preset cohesion threshold. The cohesion of each character connection position is compared with the preset cohesion threshold, and the character connection positions with a cohesion less than the preset cohesion threshold are marked as low cohesion split positions.

5. The method according to claim 1, characterized in that, The step of performing reverse segmentation on the target corpus based on the low-cohesion segmentation positions to obtain a set of candidate stable segments, and assembling a set of candidate words based on the adjacency relationships and length constraints of the set of candidate stable segments, includes: According to the low cohesion segmentation position, each text segment is segmented by breakpoint to generate the corresponding segment sequence. Segments with a length of 1 after segmentation and an occurrence frequency lower than the preset frequency threshold are eliminated, delayed, or marked as to be merged to obtain the initial stable segment sequence. Identify mixed symbol fragments containing version number, season number, region code, equipment number or character abbreviation in the initial stable fragment sequence, and perform targeted retention in combination with preset retention rules, alias normalization rules and adjacent context constraints to generate a candidate stable fragment set; Based on the sequential adjacency relationship, cross-text repetition and co-occurrence relationship, and preset length constraints of the candidate stable fragment set in the original text, restricted assembly and candidate screening are performed to obtain the candidate word set.

6. The method according to claim 1, characterized in that, The process of identifying cross-community public words based on community coverage, game coverage, and distribution dispersion, and filtering out these cross-community public words from the candidate word set, includes: Statistically analyze the frequency of occurrence of each candidate word in the candidate word set under different community identifiers and different game identifiers, and construct the community distribution matrix and game distribution matrix corresponding to the candidate words; Based on the community distribution matrix and the game distribution matrix, the community coverage, game coverage, frequency ratio difference and distribution entropy value corresponding to each candidate word are determined, and a cross-community distribution representation is formed. The cross-community distribution representation is matched with the preset public word determination conditions. Candidate words that meet the conditions of game coverage greater than a preset coverage threshold and distribution entropy value greater than a preset discrete threshold are marked as cross-community public words and filtered out from the candidate word set.

7. The method according to claim 1, characterized in that, The filtered candidate words are combined with game context trigger features and context co-occurrence relationships to perform entity confirmation and ranking, and output the entity word mining results corresponding to the target game community, including: Extract the context window of each filtered candidate word in the target corpus, identify the role, equipment, dungeon, event, version, season and gameplay trigger words associated with the candidate words, and generate the game context trigger representation corresponding to the candidate words; Based on the game context triggering representation and the co-occurrence objects, co-occurrence frequency and positional adjacency relationships in the context window, a context co-occurrence representation corresponding to the candidate word is constructed, and the entity confidence corresponding to each candidate word is determined accordingly. The filtered candidate words are sorted and selected according to the entity confidence level and preset sorting rules to generate entity word mining results corresponding to the target game community.

8. A device for mining entity words in a game community, characterized in that, include: The acquisition module is used to acquire text corpora corresponding to multiple game communities, and perform noise reduction and normalization, mixed symbol normalization, and association processing of community identifiers and game identifiers on the text corpora to generate a target corpus set; The calculation module is used to calculate the cohesion corresponding to the character connection position based on the co-occurrence statistics and edge probability distribution of adjacent characters in the target corpus, and to determine a preset cohesion threshold based on the cohesion distribution of the whole corpus, and to identify low cohesion segmentation positions based on the preset cohesion threshold. The segmentation module is used to perform reverse segmentation on the target corpus based on the low cohesion segmentation position to obtain a set of candidate stable segments, and to assemble a set of candidate words based on the adjacency relationship and length constraints of the set of candidate stable segments. The identification module is used to statistically analyze the frequency distribution of each candidate word in multiple game communities and multiple game identifiers, identify cross-community public words based on community coverage, game coverage and distribution dispersion, and filter out the cross-community public words from the candidate word set; The output module is used to perform entity confirmation and ranking on the filtered candidate words by combining game context trigger features and context co-occurrence relationships, and output the entity word mining results corresponding to the target game community.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images