Game entity word mining method and device based on structural prior and backtracking verification

Abstract

The application provides a game entity word mining method and device based on structure priori and backtracking verification. The method comprises the following steps: obtaining a corpus set corresponding to a target game community, and performing character normalization, noise cleaning and mixed expression reservation processing to obtain a target corpus; generating a continuous segment of a binary to a preset maximum length from the target corpus, calculating the internal cohesion degree of each continuous segment according to the granularity, and constructing a high-quality segment set; performing coarse-grained segmentation on the target corpus based on the high-quality segment set, and retaining a continuous segment that can be decomposed into multiple sub-segments belonging to the high-quality segment set as a candidate segment; performing backtracking verification on the candidate segment to obtain a target entity candidate word; and outputting an entity word mining result corresponding to the target game community based on the game context trigger features and the context co-occurrence relationship corresponding to the target entity candidate word. The application can improve the entity word mining accuracy, reduce the false word misrecognition rate, and enhance the incremental update stability.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for mining game entity words based on structural priors and backtracking verification. Background Technology

[0002] With the development of the online gaming industry, a large amount of text data related to characters, equipment, dungeons, events, versions, and gameplay has been continuously accumulated in game forums, community comments, strategy discussions, live stream comments, and chat texts. Accurately mining game entity words from these community corpora is a crucial foundation for achieving game content understanding, knowledge graph construction, public opinion analysis, intelligent retrieval, and operational analysis. Therefore, entity word mining for game community scenarios has become an important technical requirement in the field of game data processing.

[0003] Existing technologies typically employ methods such as dictionary matching, fixed-granularity segment statistics, mutual information calculation, left and right information entropy filtering, or general new word discovery models to identify candidate words from community texts. While these methods are applicable to general text scenarios, game community corpora contain a large number of abbreviations, slang, versioned expressions, alphanumeric mixed expressions, and temporary combinations. Furthermore, word boundaries are unstable, semantics change rapidly, and there is a lot of noise, making processing methods that rely solely on single-granularity segment statistics or static cohesion determination unsuitable for game scenarios.

[0004] In this context, existing technologies have at least the following problems: First, they are prone to over-segmenting real game entities, which can damage the entity structure; second, they are prone to mis-splitting multiple high-frequency segments into pseudo-words, resulting in insufficient accuracy of candidate results; and third, they lack a dynamic update mechanism that adapts to the incremental corpus and changes in hot topics in the community, making it difficult to balance the accuracy and stability of entity word mining. Summary of the Invention

[0005] In view of this, embodiments of this application provide a game entity word mining method and apparatus based on structural prior and backtracking verification to solve the problems of over-segmentation of real game entities, mis-spelling of pseudo-words, and insufficient stability of incremental updates in the prior art.

[0006] A first aspect of this application provides a game entity word mining method based on structural prior and backtracking verification, comprising: acquiring a corpus set corresponding to a target game community, and performing character normalization, noise removal, and mixed expression retention processing to obtain target corpus; generating continuous segments of binary to preset maximum length from the target corpus, calculating the internal cohesion of each continuous segment at granularity, and constructing a high-quality segment set by combining length binning thresholds and frequency constraints, and updating the entry and exit thresholds of the high-quality segment set based on incremental corpus; performing coarse-grained segmentation of the target corpus based on the high-quality segment set, and retaining continuous segments that can be decomposed into multiple sub-segments belonging to the high-quality segment set as candidate segments; performing backtracking verification on the candidate segments, filtering candidate segments whose length is not greater than the preset maximum length and do not belong to the high-quality segment set, and filtering candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, to obtain target entity candidate words; and outputting the entity word mining results corresponding to the target game community based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words.

[0007] A second aspect of this application provides a game entity word mining device based on structural prior and backtracking verification, comprising: an acquisition module, configured to acquire a corpus set corresponding to a target game community, and perform character normalization, noise removal, and mixed expression preservation processing to obtain the target corpus; a generation module, configured to generate continuous segments of binary to preset maximum length from the target corpus, calculate the internal cohesion of each continuous segment at the granularity, and construct a high-quality segment set by combining length binning thresholds and frequency constraints, and update the entry and exit thresholds of the high-quality segment set based on incremental corpus; and a segmentation module, configured to segment the high-quality segment set based on the high-quality segment set. The target corpus undergoes coarse-grained segmentation, retaining continuous segments that can be decomposed into multiple sub-segments belonging to a high-quality segment set as candidate segments. A filtering module performs backtracking verification on the candidate segments, filtering those whose length is not greater than a preset maximum length and do not belong to the high-quality segment set, and filtering those whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, thus obtaining target entity candidate words. An output module outputs the entity word mining results corresponding to the target game community based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words.

[0008] 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.

[0009] 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.

[0010] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring the corpus corresponding to the target game community and performing character normalization, noise removal, and mixed expression preservation processing, the target corpus is obtained. Continuous segments of binary to preset maximum length are generated from the target corpus. The internal cohesion of each continuous segment is calculated at granularity, and a high-quality segment set is constructed by combining length binning thresholds and frequency constraints. The entry and exit thresholds of the high-quality segment set are updated based on incremental corpus data. Coarse-grained segmentation is performed on the target corpus based on the high-quality segment set, retaining continuous segments that can be decomposed into multiple sub-segments belonging to the high-quality segment set as candidate segments. Backtracking verification is performed on the candidate segments, filtering out candidate segments whose length is not greater than the preset maximum length and do not belong to the high-quality segment set, and filtering out candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, thus obtaining target entity candidate words. Based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words, the entity word mining results corresponding to the target game community are output. This application can improve the accuracy of entity word mining, reduce the false word identification rate, and enhance the stability of incremental updates. Attached Figure Description

[0011] 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.

[0012] Figure 1 This is a flowchart illustrating the game entity word mining method based on structural prior and backtracking verification provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the game entity word mining device based on structural prior and backtracking verification 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

[0013] 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.

[0014] In existing technologies, entity word mining in community texts typically employs methods such as dictionary matching, fixed-granularity segment statistics, cohesion calculation, left and right information entropy filtering, or general new word discovery models. While these methods have some applicability in general text scenarios, game community corpora contain a large number of character abbreviations, equipment slang, dungeon abbreviations, versioning expressions, and mixed expressions using letters, numbers, and symbols. Furthermore, community texts are noisy, have unstable word boundaries, and evolve hotspots rapidly, making it difficult for existing technologies to simultaneously ensure entity integrity, pseudo-word filtering capabilities, and dynamic adaptability.

[0015] Based on this, existing technologies have at least the following problems: On the one hand, they are prone to over-segmenting real game entities, resulting in the destruction of entity structures such as character names, equipment names, or gameplay names; on the other hand, they are prone to mis-splitting multiple high-frequency segments into pseudo-words, resulting in insufficient accuracy of candidate results; in addition, in the face of continuously arriving incremental community corpora, existing technologies generally lack a stable dynamic update mechanism, which is prone to frequent changes in word sets due to short-term hot spot fluctuations, affecting the stability of entity word mining results.

[0016] To address the aforementioned issues, this application proposes a game entity word mining method based on structural prior and backtracking verification. The method first acquires the corpus corresponding to the target game community and performs character normalization, noise removal, and mixed expression preservation processing to obtain the target corpus. Then, it generates continuous segments of binary to preset maximum length from the target corpus, calculates the internal cohesion of each continuous segment at a granular level, and constructs a high-quality segment set by combining length binning thresholds and frequency constraints. Simultaneously, it updates the entry and exit thresholds of the high-quality segment set based on incremental corpus data. On this basis, it performs coarse-grained segmentation of the target corpus based on the high-quality segment set, retaining continuous segments that can be decomposed into multiple sub-segments belonging to the high-quality segment set as candidate segments. Next, it performs backtracking verification on the candidate segments, filtering out candidate segments that do not meet the corresponding set constraints to obtain target entity candidate words. Finally, based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words, it outputs the entity word mining results corresponding to the target game community.

[0017] Through the above technical solutions, this application can improve the accuracy of entity word mining, reduce the false recognition rate of pseudo-words, and enhance the stability of incremental updates. Specifically, by constructing a high-quality fragment set to form a structural prior, excessive segmentation of real entities can be reduced; by combining coarse-grained segmentation with backtracking verification in the processing chain, pseudo-combined fragments can be effectively filtered; by introducing an incremental update mechanism and entry and exit threshold constraints, the update stability of the high-quality fragment set under the changing community hotspot scenarios can be improved, thereby making the entity word mining results in the game community scenario more reliable and applicable.

[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 entity word mining method based on structural prior and backtracking verification provided in this application embodiment. For example... Figure 1 As shown, the method may specifically include: S101: Obtain the corpus corresponding to the target game community, and perform character normalization, noise removal and mixed expression preservation processing to obtain the target corpus; S102: Generate continuous segments of binary to preset maximum length from the target corpus, calculate the internal cohesion of each continuous segment by granularity, and construct a high-quality segment set by combining the length bucketing threshold and frequency constraints. Then, update the entry threshold and exit threshold of the high-quality segment set based on the incremental corpus. S103, based on the high-quality fragment set, performs coarse-grained segmentation on the target corpus, and retains continuous fragments that can be decomposed into multiple sub-fragments belonging to the high-quality fragment set as candidate fragments; S104, perform backtracking verification on the candidate segments, filter candidate segments whose length is not greater than the preset maximum length and do not belong to the high-quality segment set, and filter candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, to obtain the target entity candidate words; S105, based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words, outputs the entity word mining results corresponding to the target game community.

[0020] In some embodiments, a corpus corresponding to the target game community is obtained, and character normalization, noise removal, and mixed expression preservation processing are performed to obtain the target corpus, including: Collect post texts, comment texts, reply texts, chat texts, and activity discussion texts from the target game community, and associate them with time markers, topic markers, and server markers to construct a corpus; Perform full-width and half-width character unification, uppercase and lowercase character unification, duplicate character compression, emoji replacement, link cleanup, and abnormal whitespace cleanup on the corpus; Mixed representation preservation processing is performed on continuous expressions containing game names, character names, equipment names, version identifiers, numerical markers, and symbol combinations to obtain the target corpus.

[0021] Specifically, this embodiment addresses the problems of scattered sources, arbitrary expressions, frequent abbreviation mixing, and easily disrupted entity boundaries in game community corpora. It describes the process of obtaining the corpus corresponding to the target game community and performing character normalization, noise removal, and mixed expression preservation processing to obtain the target corpus. The process first aggregates multi-source texts around the same target game community and associates each text with a time identifier, topic identifier, and server identifier to form a traceable corpus. Then, it performs normalization and noise removal processing on the corpus to eliminate statistical bias caused by non-semantic differences. Finally, it performs mixed expression preservation processing on continuous expressions containing game scene-specific naming rules, thereby providing stable input for subsequent continuous segment generation, internal cohesion calculation, and the construction of high-quality segment sets.

[0022] The time identifier is used to represent the time window in which the text was generated, and can be recorded according to the posting time, reply time, event release time, or chat message generation time; the topic identifier is used to represent the discussion topic to which the text belongs, and can correspond to topic sections such as character guides, dungeon mechanics, equipment combinations, event redemption, version changes, or season discussions; the server identifier is used to represent the server, channel server, or season server source of the text. Mixed expression preservation processing means that continuous expressions containing Chinese characters, letters, numbers, or symbols are not directly separated, but are preserved as a whole in the preprocessing stage, so that structures such as game name + version identifier, character name + numerical marker, equipment name + enhancement symbol, gameplay abbreviation + mode identifier remain intact in subsequent statistics.

[0023] For example, targeting a specific action role-playing game community, we collected text from its official forum, including post text, post comments, thread replies, guild chat text, and discussion text of limited-time events. During collection, we recorded phrases like "Today's Abyss Level 7 Dragon Lancer with Starfall Bow is really strong," "I've completed my Season 3 Shadow Assassin account," "Seeking advice on boss rush strategies," and "Should I replace my 3+2 gear with the Dawnbreaker Ring?" as separate corpora, writing them into a corpus. We also simultaneously recorded the corresponding time markers, topic markers, and server markers. Posts analyzing the Abyss Level 7 mechanics were associated with the dungeon topic marker, comments recommending team compositions for Season 3 were associated with the season topic marker, and text from the official server and the test server were associated with different server markers. This way, in subsequent processing, the usage of the same term in different time windows, different topic threads, and different server environments can be recorded separately.

[0024] In the character normalization and noise cleaning stage, full-width and half-width unification is performed on the corpus set, so that Season S3 is unified as Season S3; case unification is performed, so that BossRush and BOSSRUSH are unified as bossrush; repeated character compression is performed, so that "chong chong chong chong qiang a a a" is compressed into "chong chong qiang a"; emoji replacement is performed, so that [Laugh] and [Cry] are replaced with preset emotion placeholder marks; link cleaning is performed, so that short links, jump links and meaningless URL fields in activity posts are deleted; abnormal blank cleaning is performed, so that extra line breaks, consecutive tab characters and meaningless spaces are removed. After the above processing, the superficial differences in the text are unified, while the core strings related to game entity recognition are still retained.

[0025] In the mixed expression retention processing, overall recognition and continuous retention are performed on continuous expressions containing game names, character names, equipment names, version identifiers, numerical marks and symbol combinations. For example, "Dragonlance Knight S3", "Shadow Slayer +12", "T0 lineup", "3-star bow", "PVE server", "version 2.4", and "bossrush dungeon" are identified as mixed expression candidates, no segmentation boundary is forcibly inserted between "Dragonlance Knight" and S3, or between "Shadow Slayer" and +12, and bossrush is not split into separate letter strings. For continuous expressions composed of mixed Chinese characters and numbers such as "Season Server Newbie Party Abyss Floor 7 Speedrun Team", the system first performs overall marking according to preset retention rules and context connection patterns, and then outputs them as retention units in the target corpus. Based on the target corpus, continuous fragments from bigram to a preset maximum length can be further generated subsequently, for constructing a high-quality fragment set.

[0026] Through the above embodiments, multi-source heterogeneous texts in the game community can form standardized corpus inputs with identification constraints, reduce the interference of noisy expressions and writing differences on continuous fragment statistics, and avoid the damage to game names, character names, equipment names and versioned mixed expressions in the preprocessing stage, thereby providing consistent, stable and scenario-adaptive basic corpus for subsequent game entity word mining.

[0027] In some embodiments, continuous fragments from bigram to a preset maximum length are generated for the target corpus, the internal cohesion of each continuous fragment is calculated according to granularity, a high-quality fragment set is constructed in combination with length binning thresholds and frequency constraints, and entry thresholds and exit thresholds of the high-quality fragment set are updated based on incremental corpus, which includes: Generating continuous fragments of different granularities for the target corpus by a sliding window, and counting the occurrence frequency, decomposition structure frequency and boundary co-occurrence information of each continuous fragment; Determining the internal cohesion of each continuous fragment based on the overall occurrence probability of each continuous fragment and the combined probability corresponding to the decomposition structure, and constructing the high-quality fragment set in combination with the length binning threshold corresponding to the fragment length and the minimum frequency constraint; Extract continuous segments and their decomposition structures from the incremental corpus, update the corresponding statistics, and perform incremental updates on the high-quality segment set based on the entry threshold, exit threshold, and time window stability constraints.

[0028] Specifically, this embodiment generates continuous segments of 2-gram length to a preset maximum length from the target corpus, calculates the internal cohesion of each continuous segment at a granular level, and constructs a high-quality segment set by combining length binning thresholds and frequency constraints. The implementation process of updating the entry and exit thresholds of the high-quality segment set based on incremental corpus is also described. This process is performed on the target corpus after the aforementioned character normalization, noise removal, and mixed expression preservation processing. Its purpose is to extract segments with stable internal structures from the game community corpus, providing structural priors for subsequent coarse-grained segmentation and backtracking verification.

[0029] In this embodiment, the preset maximum length can be set to 5, and the post title, body paragraphs, comment statements, chat phrases, and activity discussion fragments can be used as processing units. Each processing unit is scanned in character order using a sliding window to generate 2-ary, 3-ary, 4-ary, and 5-ary continuous fragments respectively. For example, for the aforementioned corpus, "Abyss 7th Floor Dragon Lancer Knight with Meteor Bow is really strong," 2-ary fragments such as Abyss, Abyss 7, 7th Floor, Dragon Lancer, Lancer Knight, Knight, Meteor, and Meteor Bow can be generated; 3-ary fragments such as Abyss 7, Dragon Lancer Knight, Lancer Knight, and Meteor Bow can be generated; 4-ary fragments such as Abyss 7th Floor and Dragon Lancer Knight can be generated; and 5-ary fragments such as Abyss 7th Floor Dragon and Layer Dragon Lancer Knight can be generated. For "Seeking advice on boss rush tactics," 2-ary fragments such as bo, os, ss, sr, ru, us, and sh can be generated; and 3-ary fragments such as bos, oss, ssr, sru, rus, and ush can be generated. While generating continuous segments at various granularities, the system simultaneously calculates their frequency of occurrence, frequency of decomposition structures, and co-occurrence information of boundaries. The frequency of decomposition structures records the combination of left and right sub-segments after the continuous segment is split at different points. Boundary co-occurrence information records the co-occurrence distribution of adjacent characters, symbols, or stop boundaries on the left and right sides of the segment. For example, for "Dragon Lance Knight," the system can record the combination statistics of different decomposition structures such as "Dragon Lance" and "Knight," and "Dragon Lance Knight" and "Soldier." For "Shadow Killer +12," the system can record common collocations on the left side of "Shadow Killer" and common boundaries on the right side of "+12" while preserving mixed representations, to facilitate subsequent identification of whether the segment has stable internal connectivity.

[0030] After completing the statistics, the system determines the internal cohesion based on the overall occurrence probability of each continuous segment and the combination probability corresponding to the decomposition structure. Specifically, for each segmentation position of the same continuous segment, the ratio of the overall occurrence probability to the corresponding combination probability can be calculated, and the minimum ratio is selected as the internal cohesion characterization of that continuous segment, thus reflecting the internal stability of the segment at the weakest segmentation position. For example, if the Dragonlance Knight has a high overall occurrence frequency, and although decompositions such as Dragonlance and Knight, and Dragonlance Knight and Warrior also occur, their combination probabilities are significantly lower than the overall occurrence probability, then the Dragonlance Knight can be judged to have high internal cohesion. Conversely, for combinations like the large-scale type, which are temporarily spliced ​​together from common short segments, because their overall occurrence probability is relatively low, and the combination probability at a certain segmentation position is close to or even higher than the overall occurrence probability, their internal cohesion is low.

[0031] Furthermore, length binning thresholds and minimum frequency constraints are set for different fragment lengths. For example, a higher frequency threshold is used for binary fragments, while a relatively lower but stricter cohesion threshold is used for quaternary and quinary fragments, thus balancing the differences between high-frequency noise in short fragments and low-frequency sparsity in long fragments. Continuous fragments that meet the corresponding length binning threshold and reach the minimum frequency constraint are written into the high-quality fragment set. For example, fragments such as Dragon Lance Knight, Starfall Bow, Abyss Level 7, Boss Rush, and Shadow Assassin can enter the high-quality fragment set, while combined fragments that only appear occasionally in individual texts and have insufficient cohesion are not retained.

[0032] During the incremental update phase, the system continuously receives new posts, activity discussions, and chat messages to form incremental corpora. For example, after the launch of Season 3, new expressions such as "Season 3 raid group," "Dragon Lancer S3," and "bossrush" appeared. At this time, the system does not rescan the entire historical corpus but only extracts continuous segments and their decomposition structures from the incremental corpus, updating the corresponding occurrence frequency, decomposition structure frequency, and boundary co-occurrence information. Subsequently, incremental updates are performed on the high-quality segment set based on entry thresholds, exit thresholds, and time window stability constraints.

[0033] In some examples, segments whose frequency consistently increases across multiple time windows and whose internal cohesion remains consistently above the entry threshold are added to the high-quality segment set. Segments that appear momentarily due to short-term trending terms but subsequently fall below the exit threshold or rapidly decline in subsequent time windows are either not added or removed from the set. For instance, a seasonal server raid group appearing consistently in multiple strategy topics, server discussions, and chat texts can be reliably included in the high-quality segment set; while a slogan-like combination appearing only briefly in short-term reposts and then quickly disappearing is excluded because it does not meet the time window stability constraint. This approach allows the high-quality segment set to maintain update stability while adapting to the emergence of new community terms.

[0034] Through the above implementation methods, it is possible to generate continuous segment statistical results covering different granularities in the diverse texts of the game community, accurately identify high-quality segments with stable internal structures, and combine incremental corpora to achieve dynamic maintenance of the high-quality segment set, thereby providing a stable and updatable structural prior for subsequent coarse-grained segmentation and backtracking verification, thereby improving the accuracy of game entity word recognition and enhancing the stability of word set updates.

[0035] In some embodiments, the internal cohesion of each continuous segment is determined based on the overall occurrence probability of each continuous segment and the combination probability corresponding to the decomposition structure. A high-quality segment set is then constructed by combining a length bucketing threshold corresponding to the segment length and a minimum frequency constraint, including: For each continuous segment, a corresponding decomposition structure is generated according to different segmentation positions, and the combination probability corresponding to each decomposition structure is determined respectively. Based on the overall occurrence probability and the probability of each combination, the cohesion value corresponding to each segmentation position is determined, and the internal cohesion of the continuous segment is determined according to the target value in each cohesion value. The internal cohesion is matched and compared with the length binning threshold corresponding to the fragment length of the continuous fragments. Combined with the occurrence frequency and minimum frequency constraints of the continuous fragments, a high-quality fragment set is obtained.

[0036] Specifically, this embodiment is based on the aforementioned completion of the target corpus construction and the generation of continuous segments at different granularities. Its core lies in: establishing a set of decomposition structures at multiple segmentation positions for each continuous segment, calculating the relative relationship between the overall occurrence probability and the combination probability of each decomposition structure to obtain the internal cohesion that can reflect the internal connection strength of the segment, and then performing differentiated screening according to the bucket where the segment length is located, so as to avoid short segments being mistakenly retained due to high frequency co-occurrence, and also to avoid long segments being directly excluded due to low frequency, so that the segments entering the high-quality segment set have both structural stability and statistical reliability.

[0037] Here, decomposition structure refers to the combination of left and right sub-segments obtained by splitting the same continuous segment according to different segmentation positions. For example, for Dragonlance Knight, decomposition structures such as Dragon and Lance Knight, Dragonlance and Knight, and Dragonlance Knight and Warrior can be formed; for Abyss 7, decomposition structures such as Abyss and Abyss 7 can be formed, Abyss and 7, and Abyss 7 and 7 can be formed. Combination probability refers to the statistical probability that the corresponding left and right sub-segments form a combination in the target corpus, while internal cohesion is used to measure the tightness of the continuous segment as a whole relative to its combination after being split. In this embodiment, the target value of the cohesion value corresponding to each segmentation position can be set to the minimum value so that the internal cohesion reflects the stability of the weakest connection position, avoiding the misjudgment of a continuous segment as a whole as a stable segment because it only performs well at a few segmentation positions.

[0038] For example, in an action role-playing game community, for consecutive segments such as Dragonlance Knight, Starfall Bow, Abyss Level 7, Boss Rush, and Shadow Assassin in the aforementioned target corpus, the system first extracts the overall frequency of each segment and determines the corresponding overall probability of occurrence based on the statistical results of all segments of the same granularity. Then, for each consecutive segment, a decomposition structure is generated according to all separable positions, and the occurrence of the left and right sub-segments in each decomposition structure is statistically analyzed to obtain the corresponding combination probability. Taking Dragonlance Knight as an example, if it appears continuously as a whole in multiple thematic threads such as character guides, team discussions, and seasonal server early access recommendations, while combinations such as Dragonlance and Knight, or Dragonlance Knight and Warrior also appear, but their combination probability is significantly lower than the overall probability of Dragonlance Knight's appearance, then the cohesion values ​​corresponding to each segmentation position are high, and its minimum cohesion value can still reach the threshold corresponding to high-length bins. With the minimum frequency constraint, it can enter the high-quality segment set.

[0039] For example, regarding "bossrush," although it consists of a string of letters, it appears repeatedly in strategy requests, speedrun records, and event dungeon discussions. Furthermore, the probabilities of combinations like "bo" and "ssrush," "bos" and "srush," and "boss" and "rush" obtained from different segmentation points are all lower than the overall probability of occurrence. Therefore, it can still be considered a fragment with a stable internal structure. In contrast, for combinations like "scale-based" and "fate-based," which are temporarily pieced together from common short fragments, the overall probability of occurrence is low, and the probability of combinations at certain segmentation points is close to or even higher than the overall probability of occurrence. This results in a low minimum cohesion value, and even if individual local structures are relatively common, it is difficult to meet the corresponding length bucketing threshold.

[0040] Furthermore, the system compares the internal cohesion with the length binning threshold corresponding to the fragment length. For binary fragments, a higher minimum frequency constraint is set to suppress high-frequency noise, while for quaternary and quinary fragments, a screening condition that emphasizes cohesion is applied to accommodate the sparse but structurally stronger characteristics of long fragments. Finally, only continuous fragments that simultaneously meet both the internal cohesion and minimum frequency requirements are written into the high-quality fragment set.

[0041] Through the above implementation methods, it is possible to identify structural segments with truly stable internal connectivity relationships in continuous segments of different granularities, reduce misjudgments caused by relying solely on local high-frequency co-occurrence, and improve the reliability of high-quality segment sets through the collaborative screening of length binning thresholds and minimum frequency constraints, thereby providing more accurate structural priors for subsequent coarse-grained segmentation and backtracking verification.

[0042] In some embodiments, coarse-grained segmentation is performed on the target corpus based on a high-quality fragment set, retaining consecutive fragments that can be decomposed into multiple sub-fragments belonging to the high-quality fragment set as candidate fragments, including: Construct a continuous segment search range for the target corpus in character order, and generate a plurality of continuous segments to be determined corresponding to each starting position; For each continuous segment to be determined, retrieve the corresponding segmentation path in a length-first manner, and determine whether all sub-segments in each segmentation path belong to the high-quality segment set; When a segmentation path that meets the condition exists, retain the corresponding continuous segment to be determined as a candidate segment, and continue to perform subsequent coarse-grained segmentation based on the candidate segment.

[0043] Specifically, this embodiment is implemented on the basis of completing the foregoing construction of the target corpus, generation of continuous segments, and screening of the high-quality segment set. Its core lies in: not performing fine segmentation on the target corpus directly according to the smallest semantic unit, but based on the stable internal structure represented by the high-quality segment set, preferentially retaining longer continuous structures that may correspond to real game entities or stable community expressions, and using the retained results as input for subsequent backtracking verification, so as to avoid prematurely destroying the boundaries of real entities during the segmentation stage.

[0044] Wherein, the continuous segment search range refers to a candidate coverage interval formed by extending backward from the starting position of each character in the target corpus; the continuous segments to be determined refer to a plurality of continuous character strings located within the candidate coverage interval and having a length not exceeding a preset search upper limit; the segmentation paths refer to different decomposition schemes for splitting the same continuous segment to be determined into several sub-segments. In this embodiment, the system scans the target corpus in character order, generates a plurality of continuous segments to be determined with different lengths at each starting position, and then retrieves whether there is a feasible segmentation path for each continuous segment to be determined in a length-first manner. The so-called length-first means that it is preferentially judged whether a longer continuous segment can be composed of a plurality of sub-segments belonging to the high-quality segment set. If it exists, the longer continuous segment is retained as a whole first, instead of being immediately degraded to a shorter segmentation result.

[0045] For example, for the aforementioned target corpus "Abyss Floor 7 Dragon Lance Knight with Starfall Bow is really strong", the system constructs a continuous segment search range starting from the character "Abyss", and can sequentially generate a plurality of continuous segments to be determined such as "Abyss", "Abyss 7", "Abyss Floor 7", "Abyss Floor 7 Dragon", "Abyss Floor 7 Dragon Lance Knight" and the like. If the high-quality segment set already includes "Abyss Floor 7", "Dragon Lance Knight", and "Starfall Bow", the system can retrieve a feasible segmentation path corresponding to "Abyss Floor 7 Dragon Lance Knight", that is, composed of "Abyss Floor 7" and "Dragon Lance Knight", both of which belong to the high-quality segment set, so "Abyss Floor 7 Dragon Lance Knight" is first retained as a candidate segment, and then the same processing is continued for the subsequent part "with Starfall Bow is really strong", thereby retaining "Starfall Bow" as a subsequent candidate structure.

[0046] For example, if a Shadow Assassin character has completed their Season 3 gear set, and the high-quality fragment set includes both Season 3 gear and Shadow Assassin, the system can prioritize retaining the longer, continuous structure of Season 3 Shadow Assassin starting from the "S" position. Later, backtracking will be used to determine whether the entire structure should be retained. Similarly, regarding the boss rush strategy, if the high-quality fragment set contains both "boss," "rush," and "bossrush" fragments, the system prioritizes determining if "bossrush" has a segmentation path composed of high-quality sub-fragments. If so, the entire "bossrush" fragment is retained as a candidate fragment, instead of being split into "boss" and "rush" fragments. Regarding whether to replace the 3+2 gear with the Dawnbreaker Ring, if the high-quality fragment set includes both the 3+2 gear and the Dawnbreaker Ring, the system retains the 3+2 gear when scanning for the 3+2 and retains the Dawnbreaker Ring when scanning for the Dawnbreaker, thus preventing the numerical symbol combinations and equipment names from being broken up during the initial coarse segmentation stage.

[0047] Furthermore, for continuous structures composed of multiple strong local fragments from a high-quality fragment set, but whose overall authenticity remains to be confirmed, the system also retains them initially. For example, if the target corpus contains "Tianming Challenge Strategy Guide," and the high-quality fragment set contains both "Tianming" and "Fate," then "Tianming" can be retained as a candidate fragment in the coarse-grained segmentation stage because it has a segmentation path composed of high-quality sub-fragments. This processing does not mean that "Tianming" is directly identified as an entity word, but rather that it continues to participate in subsequent filtering as an object to be backtracked for verification. In other words, the coarse-grained segmentation in this embodiment adopts a process of first retaining and then verifying, allowing both real long entities and pseudo-combined fragments to enter the candidate range at this stage, and then further filtering is performed by the subsequent backtracking mechanism.

[0048] Through the above implementation method, a coarse-grained segmentation result with the priority preservation of long structures can be generated for the target corpus based on a high-quality fragment set. This reduces the situation where character names, equipment names, dungeon names and mixed expressions are over-segmented during the segmentation stage, and provides candidate fragments with more complete coverage for subsequent backtracking verification, thereby improving the stability of entity boundary preservation and the completeness of candidate generation during the game entity word mining process.

[0049] In some embodiments, backtracking verification is performed on candidate segments, filtering candidate segments whose length is not greater than a preset maximum length and do not belong to the high-quality segment set, and filtering candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, to obtain target entity candidate words, including: Obtain the segment length corresponding to each candidate segment, and divide each candidate segment into short segments and long segments based on the segment length; For each short segment, check whether it belongs to the high-quality segment set, and filter out short segments that do not belong to the high-quality segment set; For each long segment, extract multiple consecutive sub-segments corresponding to the preset maximum length in character order, and check whether each consecutive sub-segment belongs to the high-quality segment set; When there are continuous sub-segments that do not belong to the high-quality segment set, the corresponding long segments are filtered out, and the candidate segments that pass the backtracking verification are identified as candidate words for the target entity.

[0050] Specifically, in this embodiment, the preset maximum length can be set to 5. The system first obtains the fragment length corresponding to each candidate fragment, and then divides the candidate fragments into short fragments and long fragments according to the fragment length. Among them, candidate fragments with a fragment length of no more than 5 are classified as short fragments, and candidate fragments with a fragment length greater than 5 are classified as long fragments. For short fragments, the system directly checks whether they belong to the high-quality fragment set. If a short fragment does not enter the high-quality fragment set, it means that the fragment does not meet the structural stability requirements in the multi-granularity cohesion screening stage and should be directly filtered.

[0051] For example, in the aforementioned target corpus, if short fragments such as Dragonlance Knight, Meteor Bow, Abyss Level 7, and Boss Rush are retained during the coarse-grained segmentation stage, the system will then check whether each fragment belongs to the high-quality fragment set. For Dragonlance Knight, Meteor Bow, and Boss Rush, since they have high internal cohesion across multiple thematic threads such as character guides, dungeon discussions, and strategy requests, and have already entered the high-quality fragment set, they will continue to be retained. However, short fragments that only appear occasionally in a single comment and are not included in the high-quality fragment set will be directly removed.

[0052] For long segments, the system extracts multiple consecutive sub-segments corresponding to a preset maximum length in character order, and checks each consecutive sub-segment to see if it belongs to the high-quality segment set. The corresponding long segment is retained only if all extracted consecutive sub-segments belong to the high-quality segment set; if any consecutive sub-segment does not belong to the high-quality segment set, it is determined that there are unstable connection positions within the long segment, and the corresponding long segment is filtered out. For example, for the Abyss 7th Layer Dragon Lancer Knight segment retained in the coarse-grained segmentation stage, the system can sequentially extract consecutive sub-segments of length 5, such as Abyss 7th Layer Dragon, Abyss 7th Layer Dragon Lancer, 7th Layer Dragon Lancer Knight, and Layer Dragon Lancer Knight, and determine whether these consecutive sub-segments all belong to the high-quality segment set.

[0053] If all the above-mentioned continuous sub-fragments meet the conditions, it indicates that the Dragonlance Knight of the Abyss Level 7 has a continuous and stable local structure and can continue to be retained as a candidate word for the target entity. Similarly, for the "Heavenly Fate" retained in the Heavenly Fate Challenge, although it may be temporarily retained in the coarse-grained segmentation stage because both "Heavenly Fate" and "Fate" belong to the high-quality fragment set, in the backtracking verification stage, after extracting continuous sub-fragments of length 5 or matching the preset maximum length, if continuous sub-fragments that do not belong to the high-quality fragment set are found, then "Heavenly Fate" will be filtered out entirely. Likewise, for pseudo-combined fragments such as the large-scale type formed by splicing common short fragments, the instability of the local structure can also be identified and eliminated by searching continuous sub-fragments one by one.

[0054] After completing short and long segment filtering, the system will determine the candidate segments that have been verified through backtracking as candidate words for the target entity. For example, segments such as Dragon Knight, Starfall Bow, Abyss Level 7, Boss Rush, and Shadow Assassin can be retained and input into the subsequent game context trigger feature extraction and context co-occurrence relationship analysis process.

[0055] Through the above implementation methods, further structural legality verification can be performed on the coarse-grained segmentation results. This not only avoids the accidental deletion of real game entities due to conservative segmentation strategies, but also effectively filters out pseudo-combined fragments formed by mis-splitting of local high-quality fragments, thereby improving the accuracy of target entity candidate words and the reliability of subsequent entity confirmation processing.

[0056] In some embodiments, based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words, the entity word mining results corresponding to the target game community are output, including: Extract the context window of each target entity candidate word in the target corpus, and identify the character identifier, equipment identifier, dungeon identifier, event identifier, version identifier, season identifier and gameplay identifier that co-occur with each target entity candidate word to determine the game context trigger features; Based on the co-occurrence associations of candidate words for each target entity under different topic threads, different time windows, and different forms of expression, a contextual co-occurrence relationship is constructed. Based on game context trigger features and context co-occurrence relationships, entity confirmation, alias merging and sorting are performed on each target entity candidate word, and the entity word mining results are output.

[0057] Specifically, this embodiment is carried out after the aforementioned completion of target corpus construction, continuous segment screening, coarse-grained segmentation and backtracking verification. Its purpose is to perform further semantic confirmation on the target entity candidate words that have passed the structural verification, to distinguish the candidate segments that only have structural stability from the entity words that truly have the semantic orientation of the game scene, and to merge different expressions of the same entity to form entity word mining results that can be directly used for community content understanding and knowledge organization.

[0058] Among them, game context triggering features refer to the contextual marking features formed when the target entity candidate word co-occurs with game scene elements in the target corpus, which may include character identifiers, equipment identifiers, dungeon identifiers, event identifiers, version identifiers, season identifiers, and gameplay identifiers. Context co-occurrence relationship refers to the stable association relationship formed between the target entity candidate word and other candidate words or scene identifiers under different topic threads, different time windows, and different expression forms. Entity confirmation refers to determining whether the target entity candidate word corresponds to a real game entity based on game context triggering features and context co-occurrence relationship; alias merging refers to aggregating the full name, abbreviation, nickname, and mixed spelling of the same entity appearing in the community into the same entity item; sorting processing refers to determining the output priority according to the semantic stability of entity words in the target game community, scene triggering strength, and cross-thread recurrence.

[0059] For example, in the aforementioned target corpus, the system extracted context windows from candidate words for target entities such as Dragonlance Knight, Meteor Bow, Abyss Level 7, Boss Rush, and Shadow Assassin. For Dragonlance Knight, its context window frequently contains identifiers such as character, early game progression, team composition, endgame gear, and equipment setup, and it consistently appears in different thematic threads such as seasonal server early game recommendations, character strategy discussions, and equipment combination exchanges, thus identifying obvious character-related contextual triggering characteristics. For Meteor Bow, its context window often co-occurs with content such as bow, enhancement, endgame gear, affixes, and dungeon drops, confirming its equipment-related contextual triggering characteristics. For Abyss Level 7, its context consistently contains content such as mechanics, clearing, team composition, speedrun, and boss, confirming its dungeon-related contextual triggering characteristics. For Boss Rush, it appears not only in strategy requests and event dungeon discussions but also co-occurs with expressions such as rushing dungeons, pushing levels, and speedrunning; therefore, while identifying gameplay identifiers, the system also records their co-occurrence relationships with different expression forms.

[0060] Furthermore, for expressions such as "Shadow Killer Account," "Shadow Killer Account +12," and "Killer Account," the system combines the enhancement markers, equipment terms, and character pairing information appearing in the context to determine that they refer to the same equipment entity and performs alias merging. For expressions such as "S3 Season Server," "Season Server," and "S3 Server," the system combines the time window, version discussion topic, and server identifier to determine that they belong to the same seasonal scene entity cluster. For candidate fragments that only appear briefly under a single topic and lack support from character, equipment, dungeon, event, version, season, or gameplay trigger identifiers, even if they pass the aforementioned structural filtering, they will not be prioritized for final entity word output.

[0061] Furthermore, after completing entity confirmation, alias merging, and sorting, the system outputs the entity word mining results corresponding to the target game community. For example, it classifies Dragonlance Knight into the character category, Starfall Bow and Shadow Assassin into the equipment category, Abyss 7 into the dungeon category, and Bossrush into the gameplay category, and adds its main alias and active context to each entity.

[0062] Through the above implementation methods, based on structural screening, the unique semantic triggering information and cross-scene co-occurrence relationships of the game community can be further combined to effectively confirm, merge and sort the target entity candidate words, reduce the situation of erroneous output of non-entity fragments, and improve the accuracy, completeness and scene adaptability of game entity word mining results.

[0063] 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.

[0064] Figure 2 This is a schematic diagram of the game entity word mining device based on structural prior and backtracking verification provided in an embodiment of this application. For example... Figure 2 As shown, the device includes: The acquisition module 201 is used to acquire the corpus set corresponding to the target game community, and perform character normalization, noise removal and mixed expression preservation processing to obtain the target corpus; The generation module 202 is used to generate continuous segments of binary to preset maximum length from the target corpus, calculate the internal cohesion of each continuous segment by granularity, and construct a high-quality segment set by combining the length binning threshold and frequency constraints, and perform entry threshold and exit threshold update on the high-quality segment set based on incremental corpus. The segmentation module 203 is used to perform coarse-grained segmentation on the target corpus based on a high-quality fragment set, and retain continuous fragments that can be decomposed into multiple sub-fragments belonging to the high-quality fragment set as candidate fragments; The filtering module 204 is used to perform backtracking verification on the candidate segments, filter the candidate segments whose length is not greater than the preset maximum length and do not belong to the high-quality segment set, and filter the candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, so as to obtain the target entity candidate words. Output module 205 is used to output the entity word mining results corresponding to the target game community based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words.

[0065] In some embodiments, Figure 2 The acquisition module 201 collects post text, comment text, reply text, chat text, and activity discussion text from the target game community, and associates them with time identifiers, topic identifiers, and server identifiers to construct a corpus; it performs full-width and half-width character unification, uppercase and lowercase character unification, duplicate character compression, emoticon replacement, link cleanup, and abnormal whitespace cleanup on the corpus; it performs mixed expression retention processing on continuous expressions containing game names, character names, equipment names, version identifiers, numerical markers, and symbol combinations to obtain the target corpus.

[0066] In some embodiments, Figure 2 The generation module 202 generates continuous segments of different granularities from the target corpus using a sliding window, and counts the occurrence frequency, decomposition structure frequency, and boundary co-occurrence information of each continuous segment. Based on the overall occurrence probability of each continuous segment and the combination probability corresponding to the decomposition structure, the internal cohesion of each continuous segment is determined. Combined with the length bucket threshold and minimum frequency constraint corresponding to the segment length, a high-quality segment set is constructed. The continuous segments and their decomposition structures appearing in the incremental corpus are extracted, the corresponding statistics are updated, and the high-quality segment set is incrementally updated according to the entry threshold, exit threshold, and time window stability constraint.

[0067] In some embodiments, Figure 2 The generation module 202 generates corresponding decomposition structures for each continuous segment according to different cutting positions, and determines the combination probability corresponding to each decomposition structure. Based on the overall occurrence probability and each combination probability, it determines the cohesion value corresponding to each cutting position, and determines the internal cohesion of the continuous segment according to the target value in each cohesion value. It matches and compares the internal cohesion with the length bucket threshold corresponding to the segment length of the continuous segment, and combines the occurrence frequency and minimum frequency constraint of the continuous segment to filter out a high-quality segment set.

[0068] In some embodiments, Figure 2The segmentation module 203 constructs a continuous segment search range for the target corpus according to the character order and generates multiple continuous segments to be judged corresponding to each starting position. For each continuous segment to be judged, the corresponding segmentation path is retrieved in a length-first manner, and it is determined whether the sub-segments in each segmentation path belong to the high-quality segment set. When there is a segmentation path that meets the conditions, the corresponding continuous segment to be judged is retained as a candidate segment, and subsequent coarse-grained segmentation is performed based on the candidate segment.

[0069] In some embodiments, Figure 2 The filtering module 204 obtains the segment length corresponding to each candidate segment and divides each candidate segment into short segments and long segments based on the segment length; for each short segment, it checks whether it belongs to the high-quality segment set and filters short segments that do not belong to the high-quality segment set; for each long segment, it extracts multiple continuous sub-segments corresponding to the preset maximum length in character order and checks whether each continuous sub-segment belongs to the high-quality segment set one by one; when there are continuous sub-segments that do not belong to the high-quality segment set, it filters the corresponding long segments and determines the candidate segments that pass the backtracking verification as target entity candidate words.

[0070] In some embodiments, Figure 2 The output module 205 extracts the context window of each target entity candidate word in the target corpus and identifies the character identifier, equipment identifier, dungeon identifier, event identifier, version identifier, season identifier, and gameplay identifier that co-occur with each target entity candidate word to determine the game context trigger features; based on the co-occurrence association of each target entity candidate word under different topic threads, different time windows, and different expression forms, it constructs the context co-occurrence relationship; based on the game context trigger features and the context co-occurrence relationship, it performs entity confirmation, alias merging, and sorting processing on each target entity candidate word, and outputs the entity word mining results.

[0071] 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.

[0072] 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.

[0073] 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.

[0074] 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.

[0075] 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.

[0076] 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.

[0077] 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.

[0078] 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.

[0079] 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.

[0080] 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.

[0081] 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.

[0082] 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.

[0083] 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.

[0084] 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 game entity word mining method based on structural prior and backtracking verification, characterized in that, include: Obtain the corpus corresponding to the target game community, and perform character normalization, noise removal and mixed expression preservation processing to obtain the target corpus; Generate continuous segments of binary to preset maximum length from the target corpus, calculate the internal cohesion of each continuous segment by granularity, and construct a high-quality segment set by combining length bucketing threshold and frequency constraints. Then, update the entry threshold and exit threshold of the high-quality segment set based on incremental corpus. Based on the set of high-quality segments, coarse-grained segmentation is performed on the target corpus, and continuous segments that can be decomposed into multiple sub-segments belonging to the set of high-quality segments are retained as candidate segments; Perform backtracking verification on the candidate segments, filter candidate segments whose length is not greater than the preset maximum length and do not belong to the high-quality segment set, and filter candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, to obtain target entity candidate words; Based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words, 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 the corpus corresponding to the target game community and performing character normalization, noise removal, and mixed expression preservation processing to obtain the target corpus, including: Collect post texts, comment texts, reply texts, chat texts, and activity discussion texts from the target game community, and associate them with time markers, topic markers, and server markers to construct a corpus; The corpus was subjected to full-width and half-width character unification, uppercase and lowercase character unification, duplicate character compression, emoji replacement, link cleanup, and abnormal whitespace cleanup. The target corpus is obtained by performing mixed expression preservation processing on continuous expressions containing game names, character names, equipment names, version identifiers, numerical markers, and symbol combinations.

3. The method according to claim 1, characterized in that, The process of generating continuous segments of binary length to a preset maximum length from the target corpus, calculating the internal cohesion of each continuous segment at the granularity level, constructing a high-quality segment set by combining length bucketing thresholds and frequency constraints, and updating the entry and exit thresholds of the high-quality segment set based on incremental corpus includes: The target corpus is used to generate continuous segments of different granularities using a sliding window, and the occurrence frequency, decomposition structure frequency, and boundary co-occurrence information of each continuous segment are statistically analyzed. Based on the overall occurrence probability of each continuous segment and the combination probability corresponding to the decomposition structure, the internal cohesion of each continuous segment is determined, and combined with the length bucket threshold and minimum frequency constraint corresponding to the segment length, the high-quality segment set is constructed. Extract continuous segments and their decomposition structures from the incremental corpus, update the corresponding statistics, and perform incremental updates on the high-quality segment set based on the entry threshold, exit threshold, and time window stability constraints.

4. The method according to claim 3, characterized in that, The internal cohesion of each continuous segment is determined based on the overall occurrence probability of each continuous segment and the combination probability corresponding to the decomposition structure. This cohesion is then combined with a length bucketing threshold corresponding to the segment length and a minimum frequency constraint to construct the high-quality segment set, including: For each continuous segment, a corresponding decomposition structure is generated according to different segmentation positions, and the combination probability corresponding to each decomposition structure is determined respectively. Based on the overall occurrence probability and the combination probability, the cohesion value corresponding to each segmentation position is determined, and the internal cohesion of the continuous segment is determined according to the target value in each cohesion value. The internal cohesion is matched and compared with the length binning threshold corresponding to the segment length of the continuous segment, and combined with the occurrence frequency and minimum frequency constraint of the continuous segment, to obtain the set of high-quality segments.

5. The method according to claim 1, characterized in that, The step of performing coarse-grained segmentation on the target corpus based on the high-quality segment set, retaining continuous segments that can be decomposed into multiple sub-segments belonging to the high-quality segment set as candidate segments, includes: The target corpus is constructed with continuous segment search range in character order, and multiple continuous segments to be judged are generated corresponding to each starting position; For each continuous segment to be determined, the corresponding segmentation path is retrieved in a length-first manner, and it is determined whether the sub-segments in each segmentation path belong to the set of high-quality segments. When a segmentation path that meets the conditions exists, the corresponding continuous segment to be determined is retained as the candidate segment, and subsequent coarse-grained segmentation is performed based on the candidate segment.

6. The method according to claim 1, characterized in that, The process involves performing backtracking verification on the candidate segments, filtering candidate segments whose length is not greater than the preset maximum length and do not belong to the high-quality segment set, and filtering candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, to obtain target entity candidate words, including: Obtain the segment length corresponding to each candidate segment, and use the segment length to divide each candidate segment into short segments and long segments; For each of the short segments, it is retrieved whether it belongs to the set of high-quality segments, and short segments that do not belong to the set of high-quality segments are filtered out. For each of the long segments, extract multiple consecutive sub-segments corresponding to the preset maximum length in character order, and check whether each of the consecutive sub-segments belongs to the set of high-quality segments; When there are continuous sub-fragments that do not belong to the set of high-quality fragments, the corresponding long fragments are filtered out, and the candidate fragments that pass the backtracking verification are determined as the candidate words of the target entity.

7. The method according to claim 1, characterized in that, The step of outputting entity word mining results corresponding to the target game community based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words includes: Extract the context window of each target entity candidate word in the target corpus, and identify the character identifier, equipment identifier, dungeon identifier, event identifier, version identifier, season identifier and gameplay identifier that co-occur with each target entity candidate word to determine the game context triggering features; Based on the co-occurrence associations of the candidate words of each target entity under different topic threads, different time windows, and different expression forms, the context co-occurrence relationship is constructed. Based on the game context triggering features and the context co-occurrence relationship, entity confirmation, alias merging and sorting are performed on each of the target entity candidate words, and the entity word mining results are output.

8. A game entity word mining device based on structural prior and backtracking verification, characterized in that, include: The acquisition module is used to acquire the corpus corresponding to the target game community and perform character normalization, noise removal and mixed expression preservation processing to obtain the target corpus; The generation module is used to generate continuous segments of binary to preset maximum length from the target corpus, calculate the internal cohesion of each continuous segment by granularity, and construct a high-quality segment set by combining the length binning threshold and frequency constraints, and perform entry threshold and exit threshold update on the high-quality segment set based on incremental corpus. The segmentation module is used to perform coarse-grained segmentation on the target corpus based on the high-quality segment set, and retain continuous segments that can be decomposed into multiple sub-segments belonging to the high-quality segment set as candidate segments; The filtering module is used to perform backtracking verification on the candidate segments, filter candidate segments whose length is not greater than the preset maximum length and do not belong to the high-quality segment set, and filter candidate segments whose length is greater than the preset maximum length and have sub-segments corresponding to the preset maximum length that do not belong to the high-quality segment set, to obtain target entity candidate words; The output module is used to output the entity word mining results corresponding to the target game community based on the game context triggering features and context co-occurrence relationships corresponding to the target entity candidate words.

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