Public opinion news subject identification method, system and equipment and storage medium

Abstract

The invention discloses a public opinion news subject recognition method, system and device and a storage medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: carrying out the entity matching of a news text based on a hierarchical dictionary, and obtaining an initial entity and the appearance position of the initial entity in the news text; extracting an entity context set by adopting a preset two-way extension window strategy according to the occurrence position; generating a candidate entity packet based on the initial entity and the entity context set; generating an abstract by using the large model in combination with the candidate entity package; based on the abstract and the entity context set, scoring each candidate entity through a large model to obtain a quantitative score; and screening out a core subject of the public opinion news from the candidate entity packet according to the quantitative score and a preset threshold. According to the method, efficient, accurate and explainable public opinion news core subject identification is realized by fusing hierarchical entity positioning, public opinion oriented abstract generation and a large model quantitative scoring mechanism based on dynamic cue words.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an opinion news subject identification method, system, device and storage medium. BACKGROUND

[0002] The current opinion monitoring industry is in the era of "large model + multi-modal", and the technical architecture has been relatively mature. The current mainstream opinion monitoring methods include rule-based methods, statistical learning methods and deep learning methods. Some advanced systems have begun to try to introduce pre-trained language models to improve the accuracy of entity recognition.

[0003] However, although the existing opinion monitoring technology has made great progress, there are still the following obvious defects: first, the core subject screening is difficult due to the complex entities involved in opinion news; second, the manual judgment of the main mentioned entity company is extremely low in efficiency; third, the recognition efficiency of non-key news is low; fourth, the reputation risk assessment is not accurate enough.

[0004] In summary, the existing opinion monitoring method is still difficult to efficiently and accurately identify the core subject with real opinion value when facing mass, high-noise, multi-source news data.

[0005] The above content is only used to assist in understanding the technical solutions of the present application, and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0006] The main purpose of the present application is to provide an opinion news subject identification method, system, device and storage medium, aiming to solve the technical problem of how to efficiently and accurately identify the core subject in opinion news.

[0007] To achieve the above purpose, the present application provides an opinion news subject identification method, which comprises: performing entity matching on the news text based on a hierarchical dictionary, obtaining an initial entity and the occurrence position of the initial entity in the news text; According to the occurrence position, a preset bidirectional expansion window strategy is adopted to extract an entity context set corresponding to the initial entity; Based on the initial entity and the entity context set, a candidate entity package is generated; Using a pre-trained large model, an abstract is generated in combination with the candidate entity package; Based on the abstract and the entity context set, the pre-trained large model is used to score the main description of each candidate entity in the opinion, and a quantitative score is obtained; According to the quantitative score and a preset threshold, the core subject of the opinion news is screened out from the candidate entity package.

[0008] In one embodiment, the step of performing entity matching on the news text based on a hierarchical dictionary to obtain an initial entity and the location information of the initial entity in the news text includes: Based on the hierarchical dictionary, multi-pattern matching is performed on the news text to identify the initial entity; Record the position of the initial entity in the news text, including the paragraph number and character offset.

[0009] In one embodiment, the step of extracting the entity context corresponding to the initial entity based on the occurrence position using a preset bidirectional expanding window strategy includes: Centered on the location where each initial entity appears, extend forward by a predetermined first number of characters and backward by a predetermined second number of characters to form a corresponding context fragment; The context fragments corresponding to the initial entity are merged and deduplicated to generate the entity context set of the initial entity.

[0010] In one embodiment, the step of generating a summary using a pre-trained large model and the candidate entity pack includes: Input the pre-trained large model with prompts containing the news text and entity names from the candidate entity package to generate an initial summary; The initial summary is subjected to quality verification, which includes: determining whether it contains the core layer and related layer entities in the candidate entity package, whether the key descriptions of each entity are complete, and whether the semantic consistency between the initial summary and the original text reaches a preset semantic threshold. If the quality check fails, the prompt instruction is adjusted based on the missing or incomplete entity information, and the pre-trained large model is triggered to regenerate the summary until a final summary that meets the quality check requirements is generated.

[0011] In one embodiment, the step of scoring the main descriptive accuracy of each candidate entity in public opinion using a pre-trained large model based on the summary and the entity context set to obtain a quantitative score includes: The pre-trained large model is used to perform semantic analysis on the entity context set to obtain semantic analysis results, wherein the semantic analysis results include the number of description dimensions, the context coverage ratio, and the proportion of negative information. The pre-trained large model is used to perform role positioning analysis on the summary to obtain role positioning analysis results; Based on the semantic analysis results and the role positioning analysis results, and combined with the preset scoring rules, a pre-trained large model is invoked to quantify and score each candidate entity to obtain a basic score. The base score is adjusted based on the presence of negative information and the completeness of the description to obtain the final quantitative score.

[0012] In one embodiment, the step of calculating the basic score of each candidate entity based on the semantic analysis results and the role positioning analysis results, combined with preset scoring rules, includes: The dimension score is determined based on the number of described dimensions; The coverage score is determined based on the context coverage ratio. The summary role score is determined based on whether the candidate entity is the subject and the corresponding percentage of descriptive words in the role positioning analysis results. The base score is obtained by weighting the dimension score, the coverage score, and the summary role score.

[0013] In one embodiment, before the step of calling a pre-trained large language model to quantify and score each candidate entity based on the semantic analysis results and the role positioning analysis results, combined with preset scoring rules, to obtain a basic score, the method further includes: For each candidate entity, a prompt word containing task instructions, input data, scoring rules, and output format is generated, and the prompt word is input into the large model to guide the pre-trained large model to perform semantic analysis, role positioning analysis, and basic score calculation.

[0014] Furthermore, to achieve the above objectives, this application also proposes a public opinion news subject identification system, which includes: The entity matching module is used to perform entity matching on the news text based on a hierarchical dictionary, and to obtain the initial entity and the position of the initial entity in the news text. The extraction module is used to extract the entity context set corresponding to the initial entity based on the occurrence position using a preset bidirectional expanding window strategy. The entity processing module is used to generate a candidate entity package based on the initial entity and the entity context set; The summary generation module is used to generate a summary by combining a pre-trained large model with the candidate entity package; The scoring module is used to score the main descriptive content of each candidate entity in public opinion based on the summary and the entity context set, and obtain a quantitative score by using a pre-trained large model. The filtering module is used to filter out the core subjects of public opinion news from the candidate entity package based on the quantified score and the preset threshold.

[0015] In addition, to achieve the above objectives, this application also proposes a public opinion news subject identification device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the public opinion news subject identification method described above.

[0016] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the public opinion news subject identification method described above.

[0017] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the public opinion news subject identification method described above.

[0018] This application proposes a method, system, device, and storage medium for identifying the subject of public opinion news. The method includes: performing entity matching on the news text based on a hierarchical dictionary to obtain initial entities and their occurrence positions in the news text; extracting the entity context set corresponding to the initial entity using a preset bidirectional expanding window strategy based on the occurrence positions; generating a candidate entity package based on the initial entity and the entity context set; generating a summary using a pre-trained large model combined with the candidate entity package; scoring each candidate entity using the pre-trained large model based on the summary and the entity context set to obtain a quantitative score; and selecting the core subject of the public opinion news from the candidate entity package based on the quantitative score and a preset threshold. This method achieves efficient, accurate, and interpretable identification of the core subject of public opinion news by integrating hierarchical entity localization, public opinion-oriented summary generation, and a large model quantitative scoring mechanism based on dynamic prompt words. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an embodiment of the public opinion news subject identification method of this application. Figure 2This is a flowchart illustrating Embodiment 2 of the public opinion news subject identification method in this application. Figure 3 This is a schematic diagram of the module structure of the public opinion news subject identification system in an embodiment of this application; Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the public opinion news subject identification method in the embodiments of this application.

[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0024] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0025] The main solution of this application embodiment is as follows: Entity matching is performed on the news text based on a hierarchical dictionary to obtain the initial entity and its occurrence position in the news text; based on the occurrence position, a preset bidirectional expanding window strategy is used to extract the entity context set corresponding to the initial entity; a candidate entity package is generated based on the initial entity and the entity context set; a summary is generated by combining the candidate entity package with a pre-trained large model; based on the summary and the entity context set, each candidate entity is scored by the pre-trained large model to obtain a quantitative score; based on the quantitative score and a preset threshold, the core subject of the public opinion news is selected from the candidate entity package.

[0026] In this embodiment, for ease of description, the public opinion news subject identification system will be used as the execution subject in the following description.

[0027] The existing public opinion monitoring technology has the following obvious shortcomings: 1. The existing system lacks a hierarchical screening mechanism for entity associations. It can only mechanically identify all entities but cannot distinguish between "core mentions" and "marginal mentions", leading to the problem of "subject generalization" in subsequent risk assessments.

[0028] 2. In the event of a public opinion crisis, relying on manual handling not only introduces subjective biases, but also results in a processing speed that lags far behind the speed of public opinion dissemination, causing companies to miss the window of opportunity for crisis response.

[0029] 3. Existing systems typically use a uniform processing flow to analyze all news, lacking the ability to predict the importance of news.

[0030] 4. Existing reputation risk assessment systems are mainly based on sentiment and frequency of mention, lacking in-depth analysis of the relationships between entities, making it difficult to accurately assess the potential impact of complex public opinion on corporate reputation.

[0031] This application provides a solution. First, by combining an AC state machine with a hierarchical entity dictionary of "core-related-edge" entities, it quickly matches and filters edge entities, while locating the position of each entity in the text and extracting its specific context. This reduces interference from non-critical entities at the source, laying the foundation for subsequent accurate analysis. Second, it uses a large model to generate "entity-related summaries," forcibly retaining the key descriptions and relationships of core / related entities. This avoids the problem of traditional summaries losing entity information and provides a condensed core basis for the large model's judgment. Finally, it designs customized prompts for each entity, including "input data + scoring rules + output format," driving the large model to automatically quantify the entity's descriptive level from 0 to 10 points in a four-step process: "context parsing - summary location - multi-dimensional scoring - result verification." The entire process requires no manual intervention, solving the problem of low manual efficiency and achieving accurate quantification of descriptive level. Ultimately, it quickly identifies the main corporate entities mentioned in the news.

[0032] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions. The following description uses a personal computer as an example to illustrate this embodiment and the subsequent embodiments.

[0033] Based on this, the embodiments of this application provide a method for identifying the subject of public opinion news, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the public opinion news subject identification method of this application.

[0034] In this embodiment, the public opinion news subject identification method includes steps S10 to S60: Step S10: Perform entity matching on the news text based on the hierarchical dictionary to obtain the initial entity and the position of the initial entity in the news text; It should be noted that the hierarchical dictionary refers to a multi-level classification dictionary constructed based on the strength of the relevance between entities and public opinion topics, including: a core layer, used to include key entities with high attention or high influence; a related layer, used to include secondary entities that have business, cooperation, competition or event connections with entities in the core layer; and a peripheral layer, used to include auxiliary entities with strong generality and weak relevance.

[0035] Understandably, since public opinion analysis needs to focus on key subjects and suppress noise interference, step S10 is performed. By combining the AC (Aho-Corasick Automaton) state machine with a hierarchical entity dictionary, marginal entities can be quickly matched and filtered, while the position of each entity in the text can be located. This reduces interference from non-key entities from the source and lays the foundation for subsequent accurate analysis.

[0036] In one feasible implementation, step S10 may include steps S11 to S13: Step S11: Perform multi-pattern matching on the news text based on the hierarchical dictionary to identify the initial entity; In this embodiment, the Aho-Corasick Automaton (AC automaton) is used in conjunction with the hierarchical dictionary to perform multi-pattern string matching. The hierarchical dictionary includes a core layer, a related layer, and a peripheral layer, which store entity names with different levels of public opinion relevance. The AC automaton matches all dictionary entries in parallel during a single scan, efficiently identifying all initial entities appearing in the news text.

[0037] Step S12: Record the position of the initial entity in the news text, where the position includes the paragraph number and the character offset.

[0038] After completing entity matching, the system records all occurrences of each initial entity in the news text, with the corresponding location information accurate to the paragraph number and character start and end offsets, and generates a mapping table between entities and locations. For example, for the entity "Huawei", if it appears in the news text in paragraph 2 (characters 15–18), paragraph 5 (characters 22–25), and paragraph 7 (characters 8–11), the system binds and stores the above location information with the entity's unique identifier, forming a structured "entity-location mapping table", providing a precise location basis for subsequent context extraction.

[0039] Through the above steps, the news text is subjected to multi-pattern matching based on the hierarchical dictionary to identify the initial entity; the occurrence position of the initial entity in the news text is recorded, including the paragraph number and character offset. This not only significantly improves the entity matching efficiency, but also suppresses noise interference from edge layer entities from the source, laying a data foundation for subsequent intelligent context extraction and core subject quantitative scoring.

[0040] Step S20: Based on the location of occurrence, a preset bidirectional expanding window strategy is used to extract the entity context set corresponding to the initial entity; Understandably, since public opinion analysis needs to fully capture the semantic environment of entities in news, and relying solely on the entity name itself is insufficient to reflect its behavior, attributes, and relationships, step S20, by dynamically extracting the text surrounding the entity, can effectively avoid information fragmentation, thereby achieving complete coverage and structured aggregation of the entity's description content.

[0041] In one feasible embodiment, step S20 may include steps S21-S22: Step S21: Taking the location of each initial entity as the center, extend forward by a preset first number of characters and extend backward by a preset second number of characters to form a corresponding context fragment; In this embodiment, for each matched initial entity, a bidirectional expanding window strategy is used to extract the corresponding context fragment. Specifically, taking the position of the initial entity in the news text as the center, a preset first number of characters (e.g., 50 characters) is expanded forward, and a preset second number of characters (e.g., 100 characters) are expanded backward to form the context fragment corresponding to the initial entity.

[0042] It should be noted that the number of characters extended forward and backward can be dynamically adjusted according to the overall length of the news text or the boundary of the paragraph, but the total length of the extended context fragment shall not be less than a preset minimum (e.g., 30 characters) to ensure semantic integrity.

[0043] Step S22: Merge and deduplicate the context fragments corresponding to the initial entity to generate the entity context set of the initial entity.

[0044] If an initial entity appears in multiple locations in the news text, the context fragments corresponding to each location are extracted, and these fragments are merged and redundant content is removed to finally generate the entity context set for that entity.

[0045] For example, if the entity "BOE" appears twice in a news article, the system extracts its surrounding text twice, merges and removes duplicates to obtain the following context set: "BOE provides OLED screens for Huawei's new mobile phones. These screens use the latest flexible display technology, and the yield rate has been increased to 98%." This set fully covers the core descriptive information of the entity, providing high-quality input for subsequent summary generation and quantitative scoring.

[0046] Through the above steps, this embodiment can accurately aggregate multi-source contextual information while preserving semantic coherence, significantly improving the completeness and representativeness of entity description, thereby supporting high-precision identification of public opinion subjects.

[0047] Step S30: Generate a candidate entity package based on the initial entity and the entity context set; In this embodiment, after obtaining the initial entity and its corresponding entity context set through steps S10 and S20, a structured candidate entity package is constructed based on the obtained initial entity and its corresponding entity context set. This candidate entity package is output by the AC automaton in conjunction with the subsequent processing module, providing standardized input for subsequent large-scale model-driven quantitative scoring. Each candidate package includes the entity name, its level (core layer, related layer, or edge layer), its frequency of occurrence in the news text, its corresponding entity context set, and a mapping table between the entity and its location.

[0048] Through the above structured encapsulation, this embodiment not only preserves the semantic context and spatial location information of entities, but also explicitly identifies their public opinion relevance level and occurrence intensity, thereby providing a highly cohesive and low-redundancy input data foundation for subsequent large-model-based summary generation, role positioning, and multi-dimensional quantitative scoring.

[0049] Step S40: Using a pre-trained large model and combining it with the candidate entity package, a summary is generated; It should be noted that the summary refers to an entity-related summary, which, within a limited word count (e.g., within 200 words), focuses on retaining the behavior, attributes, relationships with other entities, and key event information of each candidate entity, ensuring that the description of the core subject is not omitted.

[0050] Understandably, traditional summarization methods, which typically aim at full-text semantic compression, tend to overlook detailed descriptions of specific entities, leading to the loss of key subject information in public opinion analysis. Manual annotation, on the other hand, suffers from low efficiency and strong subjectivity. Therefore, by executing step S40, which guides the pre-trained large model (qwen3) to focus on the structured entity information in the candidate entity package, it is possible to effectively avoid the dilution or omission of key entity descriptions in general summaries, thereby achieving high-fidelity, entity-aware summary generation for public opinion analysis.

[0051] In one feasible embodiment, step S40 may include steps S41 to S443: Step S41: Input a prompt instruction containing the news text and the entity names in the candidate entity package into the pre-trained large model to generate an initial summary; In this embodiment, the system constructs a structured prompt instruction that explicitly requires the large model to generate a summary within a limited word count, focusing on preserving the behavior, attributes, and relationships between each entity in the candidate entity package and other entities. By using the full text of the news article and the candidate entity list as input context, the large model is guided to focus on key subjects and generate a preliminary entity-related summary.

[0052] Step S42, perform quality verification on the initial summary. The quality verification includes: determining whether it contains the core layer and related layer entities in the candidate entity package, whether the key descriptions of each entity are complete, and whether the semantic consistency between the initial summary and the original text reaches a preset semantic threshold. After the pre-trained large model generates an initial summary, a quality check is performed on the initial summary. This quality check includes: ① determining whether the initial summary completely contains all core and related entities in the candidate entity package; ② evaluating whether the key descriptive information (such as behavior, attributes, events, etc.) of each entity is sufficient and complete; ③ calculating the semantic consistency between the initial summary and the original news text, and determining whether it reaches a preset semantic similarity threshold (e.g., by using a vector cosine similarity ≥ 0.9).

[0053] If any of the above verification items are not met, the initial digest is deemed unqualified.

[0054] Step S43: If the quality check fails, the prompt instruction is adjusted based on the missing or incomplete entity information, and the pre-trained large model is triggered to regenerate the summary until a final summary that meets the quality check requirements is generated.

[0055] In this embodiment, the system analyzes the reasons for the verification failure, such as identifying "Entity A is not mentioned" or "Entity B's attribute description is missing," and generates enhanced prompts accordingly, such as "Please supplement the relevant description of [Entity A] and improve the information of [Entity B] regarding product performance." Subsequently, the updated prompts are input back into the large model to initiate a new round of summary generation. This process can be iteratively executed until the output summary passes all quality verification items.

[0056] Furthermore, to facilitate rapid identification of the semantic roles of entities in the summary by subsequent processing modules, the system uses a unified special tag to annotate all candidate entities in the final generated summary. For example, the summary can be presented as: "[Huawei] releases a new foldable screen phone, [BOE] provides OLED screens, the flexibility of which is 30% higher than the previous generation." This explicit annotation significantly improves the recognizability of entities in the summary, providing structured support for subsequent quantitative scoring and rule matching.

[0057] Through the above steps, this embodiment realizes a fully automated closed-loop generation and verification mechanism for public opinion-oriented summaries, effectively avoiding the omission of key entity information due to large model output deviations, and significantly improving the completeness, accuracy and business adaptability of the summaries.

[0058] Step S50: Based on the summary and the entity context set, score each candidate entity in the candidate entity package using a pre-trained large model to obtain a quantized score; Understandably, traditional public opinion analysis methods often rely on keyword frequency or simple rules to judge the importance of entities, making it difficult to comprehensively measure the semantic depth, role status, and sentiment of entities in the text, which can easily lead to bias or omission in the identification of core entities. Manual judgment is inefficient and inconsistent in standards. Therefore, by executing step S50, a structured quantitative scoring mechanism is constructed by integrating role positioning information in the summary with multi-dimensional semantic features in the context. This can effectively avoid reliance on subjective experience and the one-sidedness of a single indicator, thereby achieving an objective, fine-grained, and interpretable automated evaluation of the degree of description of candidate entities.

[0059] In one feasible embodiment, step S50 may include steps S51 to S54: Step S51: Perform semantic analysis on the entity context set using the pre-trained large model to obtain semantic analysis results; It should be noted that the semantic analysis results include the following three key indicators: Number of description dimensions: The number of semantic categories involved in the candidate entity in the context, including but not limited to public opinion-related dimensions such as product, technology, finance, cooperation, risk, and market behavior; Context coverage ratio: The percentage of the total characters in the context set corresponding to an entity relative to the total length of the original news text; Negative information percentage: The proportion of text fragments containing negative semantics in the context set, where negative semantics are identified based on a pre-set negative public opinion lexicon, which covers thousands of industry-related negative keywords.

[0060] In this embodiment, the system first performs semantic analysis on the entity context using a pre-trained large model. Specifically, it first counts the dimensions of the entity description. If the number of counted dimensions is greater than or equal to a preset threshold (e.g., 3), it is determined to be a multi-dimensional description. Then, it calculates the proportion of the entity context to the total length of the news text (denoted as L%). If this L% is greater than a preset proportion threshold, it is determined to be a high-coverage description. Finally, it identifies negative words (based on a negative public opinion word library containing 5000+ negative words) and counts the proportion of negative descriptions (denoted as N%).

[0061] For example, the contextual analysis results for "Huawei" are: "4 description dimensions (product, technology, price, launch time), L%=45%, N%=0%". Step S52: Perform role positioning analysis on the summary using the pre-trained large model to obtain the role positioning analysis results; In this embodiment, the system utilizes a pre-trained large model to perform structured parsing on the aforementioned generated entity-related summaries to determine the semantic roles and related positions of each candidate entity within the summary. Specifically, firstly, based on subject-verb-object syntactic structure analysis, it determines whether the entity serves as the subject of one or more sentences in the summary; then, it counts the proportion of text characters directly describing the entity in the summary to the total length of the summary, denoted as S%; finally, it identifies the semantic relationship types between the entity and other candidate entities, such as predefined relationship patterns like "core entity – supplier," "entity – regulatory body," and "partner – technology provider."

[0062] For example, the role positioning analysis result for the entity "Huawei" can be expressed as: "Subject of the summary, S%=60%, related supplier BOE, is the core subject." This result indicates that "Huawei" occupies a dominant narrative position in the summary and has a clear supply chain relationship with "BOE."

[0063] Through the above role positioning analysis, the system can characterize the structural position of entities in public opinion events from three dimensions: narrative focus, information weight, and relationship network, providing key criteria for subsequent quantitative scoring.

[0064] Step S53: Based on the semantic analysis results and the role positioning analysis results, and in conjunction with the preset scoring rules, call the pre-trained large model to quantify and score each candidate entity to obtain a basic score. In this implementation, a four-dimensional scoring model for identifying public opinion subjects is constructed by combining semantic analysis and role positioning analysis results. This model includes descriptive dimension, context coverage, summary role, and negative information impact. The base score is determined by a combination of the descriptive dimension score, context coverage score, and summary role score, and is used to characterize the core descriptive strength of candidate entities in news reports. This quantification mechanism effectively distinguishes between core subjects, related parties, and peripheral mentions, significantly improving the objectivity and reproducibility of subject identification.

[0065] Furthermore, step S53 may also include steps S531 to S534: Step S531: Determine the dimension score based on the number of description dimensions; Specifically, if an entity involves one descriptive dimension, the dimension score (D) is 1 point; two dimensions correspond to 2 points; and three or more dimensions are uniformly assigned 3 points to reflect the saturation effect of its descriptive breadth.

[0066] Step S532: Determine the coverage score based on the context coverage ratio; For example, when the context coverage ratio L%≥60%, the coverage score is 3 points; when 40%≤L%<60%, it is 2 points; when 20%≤L%<40%, it is 1 point; and when L%<20%, it is 0 points, reflecting the contribution of context information density to the importance of the subject.

[0067] Step S533: Determine the summary role score based on whether the candidate entity in the role positioning analysis result is the subject and the corresponding percentage of descriptive words. For example, if an entity is the core subject of the summary and its description accounts for S% ≥ 50% of the total word count, the summary role score (R) is 4 points; if it is the core subject but S% < 50%, it gets 3 points; if it is not the subject but has a substantial description, it gets 1 point; if it is simply mentioned by name without any description, it gets 0 points.

[0068] Step S534: The dimension score, the coverage score, and the summary role score are weighted to obtain the base score.

[0069] In this embodiment, the base score is calculated using a linear summation method, namely: Base score = Description dimension score + Context coverage score + Summary role score.

[0070] For example, for the entity "Huawei", if its description dimensions are 4 (dimension score = 3), the context coverage ratio is 45% (coverage score = 2), and the abstract contains the core subject with S% = 60% (abstract role score = 4), then its base score is 3 + 2 + 4 = 9 points.

[0071] Through the aforementioned multi-dimensional quantitative scoring mechanism, a structured and interpretable assessment of the strength of candidate entity descriptions is achieved, providing a reliable basis for the accurate identification of public opinion news subjects.

[0072] Step S54: Adjust the base score based on the presence of negative information and the completeness of the description to obtain the final quantified score.

[0073] After obtaining the base scores for each candidate entity, the system further combines the percentage of negative information (denoted as N%) with preset multi-level scoring rules to dynamically adjust the base scores, generating a final quantified score within the range of 0-10. The percentage of negative information is calculated based on the negative description fragments in the aforementioned semantic analysis results and corresponds to a negative score (N), used to assist rule matching. For example: When N% ≥ 50%, the negative score is 4. 3 points are awarded when 30% ≤ N% < 50%; 2 points are awarded when 10% ≤ N% < 30%; When N% < 10% (i.e., there is basically no negative information), a score of 0 is awarded.

[0074] It should be noted that the negative score is not directly added to the base score, but is used as one of the key criteria for rule matching.

[0075] After calculating the base score, the base score is dynamically adjusted according to the scoring rules. Specifically, if the total base score is ≥8 and there is no significant negative information (N%<10%), the entity is determined to be the core description object throughout the document, matching rule 2, and the final quantitative score is adjusted to 10 points; if the total base score is ≥5 and there is moderate or higher negative information (N%≥30%), matching rule 3, the final quantitative score is determined within the range (which can be further refined according to the severity of the negative information); if the total base score is between 2 and 4 and there is no negative information, matching rule 4, the original base score is retained as the final quantitative score (i.e., 2-4 points); if the total base score is 0 (i.e., no effective description is obtained in the context or abstract), matching rule 5, the lowest score of 1 point is assigned, indicating that it is only nominally mentioned.

[0076] For example, for the entity "Huawei", its dimensional score D=3 (involving 4 descriptive dimensions), coverage score C=2 (contextual coverage ratio of 45%), and summary role score R=4 (it is the core subject and accounts for 60% of the descriptive word count), with a total base score of 9; after analysis, its negative information ratio N%=0%, indicating no negative content. Therefore, this entity matches rule 2, and the final quantitative score is adjusted by 10 points.

[0077] Through the rule-driven score adjustment mechanism described above, the system can maintain the objectivity of the scoring while flexibly responding to typical scenarios in public opinion events, such as positive focus, negative emphasis, weak description, or lack of substantive information, thereby achieving a more logical determination of the importance of the subject.

[0078] Step S60: Based on the quantified score and the preset threshold, the core subject of public opinion news is selected from the candidate entity package.

[0079] In this embodiment, after obtaining the final quantized score of each candidate entity, the system further performs automatic verification of the scoring results using a pre-trained large model to ensure the reliability and consistency of the scoring process. Specifically, it determines whether the quantized score conforms to a preset valid range (e.g., 0-10 points); verifies whether the context fragments, summary roles, and rule matching logic referenced in the scoring criteria are consistent with the original input data; and verifies whether the context fragments, summary roles, and rule matching logic referenced in the scoring criteria are consistent with the original input data.

[0080] After successful verification, the system outputs a structured scoring result in the format of "Entity Name-Quantitative Score-Scoring Basis" and persists it to the entity quantitative scoring library for subsequent analysis, auditing, or model iteration and optimization.

[0081] Subsequently, the system filters the quantitative scores based on preset thresholds. Under the default configuration, candidate entities with a quantitative score ≥ 5 are identified as primary mentioned entities, i.e., the core subjects of public opinion news; the remaining entities are considered secondary or peripheral mentioned objects. This threshold can be dynamically adjusted according to different public opinion scenarios (such as highly sensitive events, industry monitoring, etc.) to adapt to diverse business needs.

[0082] Through the above verification and screening mechanisms, not only is the accuracy and interpretability of the scoring results guaranteed, but the core entities with real public opinion value are also efficiently and automatically identified from a massive number of candidate entities.

[0083] The method described above performs entity matching on news text based on a hierarchical dictionary to obtain initial entities and their locations within the text. Based on these locations, a pre-defined bidirectional expanding window strategy is used to extract the entity context set corresponding to the initial entity. A candidate entity package is generated based on the initial entity and the entity context set. A summary is generated using a pre-trained large-scale model combined with the candidate entity packages. Based on the summary and the entity context set, each candidate entity is scored using the pre-trained large-scale model to obtain a quantitative score. Finally, based on the quantitative score and a pre-defined threshold, the core subject of the public opinion news is selected from the candidate entity package. This method achieves efficient, accurate, and interpretable identification of the core subject of public opinion news by integrating hierarchical entity localization, public opinion-oriented summary generation, and a large-scale model quantitative scoring mechanism based on dynamic prompt words.

[0084] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Before step S53, the method for identifying the subject of public opinion news further includes step S053: Step S053: For each candidate entity, generate prompt words containing task instructions, input data, scoring rules, and output format, and input the prompt words into the pre-trained large model to guide the pre-trained large model to perform semantic analysis, role positioning analysis, and basic score calculation.

[0085] Compared to the first embodiment, this embodiment also proposes a mechanism for accurately scoring the degree of description of public opinion entities based on multi-level customized prompts and a four-step quantitative workflow. This mechanism can significantly improve the accuracy, interpretability, and consistency of large models in judging the primary and secondary status of different entities in complex news contexts.

[0086] Specifically, this solution designs structured, customized prompts using a four-module approach: "task instructions, input data, scoring rules, and output format." It also dynamically generates unique prompts for each candidate entity, guiding the large model to perform refined semantic analysis and quantitative scoring.

[0087] The prompt first clarifies the task instructions: the model, as an expert in analyzing public opinion and news entities, is required to determine the degree of description of the target entity in the news based on the provided input data, and to give a quantitative score of 0-10 according to the scoring rules, while explaining the scoring basis and closely linking it to the context and summary details.

[0088] The input data consists of four parts: the target entity name, all context fragments of the entity extracted from the AC state set, a relational summary containing the entity's annotations, and key information from the full text.

[0089] The scoring rules are strictly limited to five points: If an entity is not mentioned at all in the text, the score is 0; if the entity is the core description object throughout the text (covering more than 60% of the text in the context, and the abstract focuses on the entity), the score is 10; if the entity description is negative (the context contains negative words such as "problem," "malfunction," "complaint," etc., and the negative content accounts for no less than 50%), the score is between 5 and 10, with severe negative being 8-10 points and general negative being 5-7 points; if the entity has only a simple description and no negative information (the context only involves 1-2 attributes or behaviors), the score is between 2 and 4, with richer descriptions resulting in higher scores; if only the entity name appears without any additional information (no description in the context, and the abstract only includes a label), the score is 1.

[0090] The output format is standardized to include three parts: target entity, quantitative score (a specific value from 0 to 10), and scoring criteria. The scoring criteria must include three aspects: contextual analysis (e.g., "The context contains 3 paragraphs of description, covering product release, price and chip configuration, with a total length of 280 characters and no negative information"), summary analysis (e.g., "The summary focuses on this entity, mentions the product type and associates with supplier information"), and rule matching (e.g., "It meets rule 2 (the core description throughout) and has no negative information, so the overall score is 8").

[0091] Through the methods described above, this embodiment dynamically embeds the specific context and summary of each entity, ensuring that the analysis focuses on the actual context of the entity and avoids cross-entity interference. At the same time, it transforms the scoring rules into clear and actionable judgment criteria, effectively suppressing subjective bias in the evaluation process of large models. In addition, the forced output of structured evidence not only enhances the interpretability of the results but also facilitates manual review and subsequent iterative optimization of the model.

[0092] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the public opinion news subject identification method of this application. Any simple transformations based on this technical concept are all within the protection scope of this application.

[0093] This application also provides a public opinion news subject identification system; please refer to [reference needed]. Figure 3 The public opinion news subject identification system includes: Entity matching module 10 is used to perform entity matching on news text based on a hierarchical dictionary, and to obtain an initial entity and the position of the initial entity in the news text; Extraction module 20 is used to extract the entity context set corresponding to the initial entity based on the occurrence position using a preset bidirectional expanding window strategy; Entity processing module 30 is used to generate a candidate entity package based on the initial entity and the entity context set; The summary generation module 40 is used to generate a summary by combining the candidate entity package with a pre-trained large model; The scoring module 50 is used to score each candidate entity in the candidate entity package based on the summary and the entity context set, using a pre-trained large model, to obtain a quantitative score. The filtering module 60 is used to filter out the core subjects of public opinion news from the candidate entity package based on the quantified score and the preset threshold.

[0094] The public opinion news subject identification system provided in this application, employing the public opinion news subject identification method in the above embodiments, can solve the technical problem of how to efficiently and accurately identify the core subjects in public opinion news. Compared with the prior art, the beneficial effects of the public opinion news subject identification system provided in this application are the same as those of the public opinion news subject identification method provided in the above embodiments, and other technical features of the public opinion news subject identification system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0095] This application provides a public opinion news subject identification device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the public opinion news subject identification method in the above embodiment 1.

[0096] The following is for reference. Figure 4The diagram illustrates a structural schematic suitable for implementing the public opinion news subject identification device in the embodiments of this application. The public opinion news subject identification device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 4 The illustrated public opinion news subject identification device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0097] like Figure 4 As shown, the public opinion news subject identification device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the public opinion news subject identification device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the public opinion news subject identification device to communicate wirelessly or wiredly with other devices to exchange data. Although the figures show public opinion news subject identification devices with various systems, it should be understood that implementing or possessing all of the systems shown is not required. More or fewer systems may be implemented alternatively.

[0098] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0099] The public opinion news subject identification device provided in this application, employing the public opinion news subject identification method in the above embodiments, can solve the technical problem of how to efficiently and accurately identify the core subjects in public opinion news. Compared with the prior art, the beneficial effects of the public opinion news subject identification device provided in this application are the same as those of the public opinion news subject identification method provided in the above embodiments, and other technical features in this public opinion news subject identification device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0100] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0102] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the public opinion news subject identification method in the above embodiments.

[0103] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0104] The aforementioned computer-readable storage medium may be included in the public opinion news subject identification device; or it may exist independently and not be assembled into the public opinion news subject identification device.

[0105] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the public opinion news subject identification device, the public opinion news subject identification device performs entity matching on the news text based on a hierarchical dictionary to obtain initial entities and their occurrence positions in the news text; extracts the entity context set corresponding to the initial entity using a preset bidirectional expanding window strategy based on the occurrence positions; generates a candidate entity package based on the initial entity and the entity context set; generates a summary using a pre-trained large model combined with the candidate entity package; scores each candidate entity based on the summary and the entity context set using the pre-trained large model to obtain a quantized score; and selects the core subject of the public opinion news from the candidate entity package based on the quantized score and a preset threshold.

[0106] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0107] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0108] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0109] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described method for identifying the subject of public opinion news. This method can solve the technical problem of how to efficiently and accurately identify the core subject in public opinion news. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the public opinion news subject identification method provided in the above embodiments, and will not be elaborated upon here.

[0110] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for identifying the subject of public opinion news.

[0111] The computer program product provided in this application can solve the technical problem of how to efficiently and accurately identify the core subjects in public opinion news. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the public opinion news subject identification method provided in the above embodiments, and will not be repeated here.

[0112] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying the subject of public opinion news, characterized in that, The method for identifying the subject of public opinion news includes: Entity matching is performed on the news text based on a hierarchical dictionary to obtain the initial entity and the position of the initial entity in the news text. Based on the location of occurrence, a preset bidirectional expanding window strategy is used to extract the entity context set corresponding to the initial entity; Based on the initial entity and the entity context set, a candidate entity package is generated; Using a pre-trained large model and combining it with the candidate entity package, a summary is generated; Based on the summary and the entity context set, a pre-trained large model is used to score each candidate entity in the candidate entity package to obtain a quantified score; Based on the quantified score and the preset threshold, the core subject of public opinion news is selected from the candidate entity package.

2. The method for identifying the subject of public opinion news as described in claim 1, characterized in that, The step of performing entity matching on the news text based on a hierarchical dictionary to obtain an initial entity and the location information of the initial entity in the news text includes: Based on the hierarchical dictionary, multi-pattern matching is performed on the news text to identify the initial entity; Record the position of the initial entity in the news text, including the paragraph number and character offset.

3. The method for identifying the subject of public opinion news as described in claim 2, characterized in that, The step of extracting the entity context corresponding to the initial entity based on the occurrence position using a preset bidirectional expanding window strategy includes: Centered on the location where each initial entity appears, extend forward by a predetermined first number of characters and backward by a predetermined second number of characters to form a corresponding context fragment; The context fragments corresponding to the initial entity are merged and deduplicated to generate the entity context set of the initial entity.

4. The method for identifying the subject of public opinion news as described in claim 1, characterized in that, The step of generating a summary using a pre-trained large model and the candidate entity package includes: Input the pre-trained large model with prompts containing the news text and entity names from the candidate entity package to generate an initial summary; The initial summary is subjected to quality verification, which includes: determining whether the initial summary contains the core layer and related layer entities in the candidate entity package, whether the key descriptions of each entity are complete, and whether the semantic consistency between the initial summary and the original text reaches a preset semantic threshold. If the quality check fails, the prompt instruction is adjusted based on the missing or incomplete entity information, and the pre-trained large model is triggered to regenerate the summary until a final summary that meets the quality check requirements is generated.

5. The method for identifying the subject of public opinion news as described in claim 1, characterized in that, The step of scoring each candidate entity in the candidate entity bag based on the summary and the entity context set, and obtaining a quantified score by using a pre-trained large model, includes: The pre-trained large model is used to perform semantic analysis on the entity context set to obtain semantic analysis results, wherein the semantic analysis results include the number of description dimensions, the context coverage ratio, and the proportion of negative information. The pre-trained large model is used to perform role positioning analysis on the summary to obtain role positioning analysis results; Based on the semantic analysis results and the role positioning analysis results, and combined with the preset scoring rules, a pre-trained large model is invoked to quantify and score each candidate entity to obtain a basic score. The base score is adjusted based on the presence of negative information and the completeness of the description to obtain the final quantitative score.

6. The method for identifying the subject of public opinion news as described in claim 5, characterized in that, The step of calculating the basic score of each candidate entity based on the semantic analysis results and the role positioning analysis results, combined with preset scoring rules, includes: The dimension score is determined based on the number of described dimensions; The coverage score is determined based on the context coverage ratio. The summary role score is determined based on whether the candidate entity is the subject and the corresponding percentage of descriptive words in the role positioning analysis results. The base score is obtained by weighting the dimension score, the coverage score, and the summary role score.

7. The method for identifying the subject of public opinion news as described in claim 5, characterized in that, Before the step of using a pre-trained large language model to quantify and score each candidate entity based on the semantic analysis results and the role positioning analysis results, combined with preset scoring rules, to obtain a basic score, the method further includes: For each candidate entity, a prompt word containing task instructions, input data, scoring rules, and output format is generated, and the prompt word is input into the pre-trained large model to guide the pre-trained large model to perform semantic analysis, role positioning analysis, and basic score calculation.

8. A public opinion news subject identification system, characterized in that, The public opinion news subject identification system includes: The entity matching module is used to perform entity matching on the news text based on a hierarchical dictionary, and to obtain the initial entity and the position of the initial entity in the news text. The extraction module is used to extract the entity context set corresponding to the initial entity based on the occurrence position using a preset bidirectional expanding window strategy. The entity processing module is used to generate a candidate entity package based on the initial entity and the entity context set; The summary generation module is used to generate a summary by combining a pre-trained large model with the candidate entity package; The scoring module is used to score each candidate entity in the candidate entity package based on the summary and the entity context set, using a pre-trained large model, to obtain a quantitative score. The filtering module is used to filter out the core subjects of public opinion news from the candidate entity package based on the quantified score and the preset threshold.

9. A device for identifying the subject of public opinion news, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the public opinion news subject identification method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the public opinion news subject identification method as described in any one of claims 1 to 7.

Images