Tourism public opinion event identification method and system based on dynamic weight adjustment

Abstract

The application discloses a tourism public opinion event identification method and system based on dynamic weight adjustment, relates to the field of tourism public opinion monitoring, and comprises the following steps: cleaning and structuring cross-platform public opinion content and its time, source, publisher and geographical location and other associated elements, mapping the text / media into a unified dimension feature representation, and forming a comparable public opinion record set; on the basis, multi-dimensional key evaluation features are extracted, a feature weight vector corresponding to a monitoring period is used for weighted scoring and threshold determination to generate an early warning and event identification; subsequently, actual evolution feedback is introduced, an identification accuracy is calculated according to a warning deviation, and the contribution degree of each feature is quantified, and then the feature weight of the next period is dynamically updated. Therefore, the scoring basis is adaptively adjusted with the evolution of public opinion, so that the identification stability and early warning reliability under continuous monitoring are improved.

Description

Technical Field

[0001] This application relates to the field of tourism public opinion monitoring technology, and in particular to a method and system for identifying tourism public opinion events based on dynamic weight adjustment. Background Technology

[0002] Tourism public opinion monitoring is the process of long-term, real-time tracking and analysis of tourism-related public opinion from public data sources such as social media, news reports, and review websites. Its purpose is to promptly identify hot topics and potential crises affecting the reputation of tourist attractions or destinations and to assist managers in formulating response strategies. With the widespread adoption of mobile internet and social media, the amount of text, images, and videos posted by tourists has exploded, providing a rich data foundation for tourism public opinion monitoring. However, this also brings challenges such as data heterogeneity, difficulties in sentiment recognition, cross-platform integration, and noise interference.

[0003] Currently, most public opinion analysis methods employ static, fixed-weight models. This means that when calculating event risk scores, the weights of characteristics such as sentiment, influence, and dissemination speed are pre-set and constant. However, public opinion evolution is a dynamic process. The importance of different characteristics to the qualitative assessment of an event varies significantly in its initial, fermentation, and decline phases (e.g., initial focus on increased attention versus later focus on the credibility of information sources). Static evaluation mechanisms prevent the system from keeping pace with the dynamic changes in public opinion composition. This makes it difficult for fixed-weight reassessment to balance early-stage sensitive identification with later-stage rumor / noise suppression, hindering the construction of accurate event propagation chains and thus limiting the accuracy of identification and the reliability of early warnings. Summary of the Invention

[0004] This application provides a method, system, storage medium, computer program product, and electronic device for identifying tourism public opinion events based on dynamic weight adjustment, which at least solves the problem that the risk assessment of tourism public opinion events in the current related technologies is difficult to adapt to the evolution of public opinion, resulting in insufficient accuracy and stability of early warning.

[0005] In a first aspect, embodiments of this application provide a method for identifying tourism public opinion events based on dynamic weight adjustment. The method includes: acquiring multi-source tourism public opinion data associated with a target tourist destination, wherein the multi-source tourism public opinion data includes at least public opinion content data and publication time information, data source identification information, publisher information, and geographical location information associated with the public opinion content data, and the public opinion content data includes text content and / or media content; performing cleaning and structuring processing on the multi-source tourism public opinion data, and mapping the public opinion content data to a unified-dimensional structured feature representation to generate a public opinion record set, wherein each public opinion record in the public opinion record set contains the structured feature representation and the corresponding publication time information, data source identification information, publisher information, and geographical location information; and extracting multi-dimensional key evaluation features for event identification from each public opinion record in the public opinion record set based on a preset feature definition to generate corresponding key evaluation features. The process involves: evaluating a set of feature vectors; obtaining the feature weight vector corresponding to the current monitoring period; performing a weighted combination process on the set of key evaluation feature vectors based on the feature weight vectors to obtain the event score result for each public opinion record; comparing the event score result with an event judgment threshold to identify potential tourism public opinion events; generating early warning information and recording the event identifier associated with the early warning information if the potential tourism public opinion event is identified; collecting actual event evolution feedback data corresponding to the event identifier after the early warning, and calculating the identification accuracy rate based on the deviation between the actual event evolution feedback data and the early warning information; calculating the contribution of each key evaluation feature based on the identification accuracy rate, and dynamically updating the feature weight vector according to the contribution to obtain the updated feature weight vector for the next monitoring period; wherein, the key evaluation feature with a larger contribution is assigned a higher weight proportion in the updated feature weight vector.

[0006] Secondly, embodiments of this application provide a tourism public opinion event identification system based on dynamic weight adjustment. The system includes: a data acquisition unit, used to acquire multi-source tourism public opinion data associated with a target tourist destination. The multi-source tourism public opinion data includes at least public opinion content data and publication time information, data source identification information, publisher information, and geographical location information associated with the public opinion content data. The public opinion content data includes text content and / or media content. A data preprocessing and structured representation unit, used to perform cleaning and structured processing on the multi-source tourism public opinion data, and map the public opinion content data into a unified-dimensional structured feature representation to generate a public opinion record set. Each public opinion record in the public opinion record set contains the structured feature representation and the corresponding publication time information, data source identification information, publisher information, and geographical location information. A key evaluation feature extraction unit, used to extract multi-dimensional key evaluation features for event identification from each public opinion record in the public opinion record set based on a preset feature definition, to generate corresponding key evaluation features. The system comprises: a feature vector set; an event scoring and early warning unit, used to obtain the feature weight vector corresponding to the current monitoring period, and perform weighted combination processing on the key evaluation feature vector set based on the feature weight vector to obtain the event scoring result of each public opinion record; comparing the event scoring result with the event judgment threshold to identify potential tourism public opinion events, and generating early warning information and recording the event identifier associated with the early warning information if the existence of the potential tourism public opinion event is determined; a feedback collection and accuracy evaluation unit, used to collect actual event evolution feedback data corresponding to the event identifier after the early warning, and calculate the identification accuracy based on the deviation between the actual event evolution feedback data and the early warning information; and a contribution calculation and weight update unit, used to calculate the contribution of each key evaluation feature based on the identification accuracy, and dynamically update the feature weight vector according to the contribution to obtain the updated feature weight vector corresponding to the next monitoring period; wherein, the key evaluation feature with a larger contribution is assigned a higher weight proportion in the updated feature weight vector.

[0007] Thirdly, an electronic device is provided, comprising: 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, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the tourism public opinion event identification method based on dynamic weight adjustment according to any embodiment of this application.

[0008] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the tourism public opinion event identification method based on dynamic weight adjustment according to any embodiment of this application.

[0009] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the tourism public opinion event identification method based on dynamic weight adjustment according to any embodiment of this application.

[0010] The tourism public opinion event identification method and system based on dynamic weight adjustment provided in this application can achieve at least the following technical effects: (1) Based on a unified structured feature space, text content and / or media content from different data sources, along with their publication time, data source identifier, publisher information, geographic location information, and other related elements, are cleaned, structured, and mapped into a unified dimension of feature representation. Then, a set of computable key evaluation feature vectors is formed within this isomorphic feature space. As a result, subsequent event scoring is no longer limited by the differences in field granularity, expression methods, and content forms of various platforms, but can instead conduct comparability evaluation of public opinion records under a unified feature scale, thereby improving the consistency and stability of event scoring criteria under cross-platform integration conditions.

[0011] (2) A closed-loop self-correction mechanism of "early warning-feedback-contribution-weight update" is constructed, making the feature weight vector a parameter system that can dynamically evolve with the monitoring cycle. This closed loop forms a measurable identification accuracy rate through feedback from the actual event evolution after the early warning and the early warning deviation. The accuracy rate is further decomposed into the contribution attribution of each key evaluation feature, so that the contribution rate is used as the basis for weight update, enabling the system to automatically complete the redistribution and recalibration of the score composition during continuous operation. As a result, the sensitivity of the event score to public opinion signals of different stages and forms can be adaptively adjusted over time, avoiding the score mismatch and drift accumulation caused by long-term solidification of weights, and enhancing the traceability and interpretability between the early warning results and subsequent evolution.

[0012] This technical solution enables the periodic adaptation of the event recognition model through contribution-driven dynamic weight updates, giving the risk scoring mechanism the ability to continuously correct itself in response to real-world evolutionary feedback. With the support of a unified structured feature space, this adaptive mechanism can stably operate on multi-source heterogeneous public opinion records, thereby maintaining the reliability and consistency of event recognition and early warning output in long-term monitoring scenarios. Attached Figure Description

[0013] 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, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating an example of a tourism public opinion event identification method based on dynamic weight adjustment according to an embodiment of this application is shown; Figure 2 A flowchart illustrating the operational mechanism of an example of a tourism public opinion event identification method based on dynamic weight adjustment according to an embodiment of this application is shown. Figure 3 The graph shows the comparison of F1 scores over consecutive evaluation periods according to different methods. Figure 4 The diagram shows the dynamic evolution curves of the weights of the five-dimensional key evaluation features according to embodiments of this application over a continuous evaluation period; Figure 5 A structural block diagram of an example of a tourism public opinion event identification system based on dynamic weight adjustment according to an embodiment of this application is shown. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] It should be noted that tourism public opinion data exhibits significant multimodal and strongly unstructured characteristics: in addition to text, the proportion of media content such as images, short videos, and live stream clips continues to rise, often accompanied by composite information such as text overlays, colloquial subtitles, voice expressions, and background music; simultaneously, tourism public opinion is often naturally tied to geographical location, and spatiotemporal information such as location accuracy, distance from scenic spots, and cross-regional diffusion trajectories are of great value in judging the correlation and scope of event dissemination. Although current technologies have begun to attempt to extract processable information from media content through OCR, speech recognition, and other means, and to mitigate semantic drift caused by regional expression differences through dialect dictionaries or error correction models, issues such as inconsistent multimodal feature scales, modality loss, and noise modal interference still easily lead to problems such as emotion recognition bias and unstable event clustering, resulting in limited overall accuracy of cross-platform fusion.

[0017] On the other hand, many current systems still tend to use fixed weights or hard-coded thresholds for event risk assessment. This involves linearly combining features such as emotional intensity, dissemination influence, and dissemination speed with preset weights to obtain a risk score, which is then compared with a threshold to trigger an early warning. Related research and engineering practices both indicate that tourism public opinion exhibits a clear phased evolution: in the early stages of an event, the rate of increase and the degree of spatiotemporal concentration often better characterize "outbreak signs"; in the fermentation and diffusion stages, dissemination influence and cross-platform reposting relationships better reflect the potential impact; and in the later stages, the credibility of the information source, noise propagation, and the infiltration of rumors become key factors affecting the reliability of early warnings. Fixed weight mechanisms are difficult to adaptively adjust with changes in stages, easily leading to a structural contradiction of "insensitivity in the early stages and high false alarms in the later stages," making it difficult for early warning systems to balance sensitivity and robustness.

[0018] Furthermore, regarding the understanding of public opinion evolution and the construction of the dissemination chain, current technologies mostly employ methods such as topic detection and tracking, keyword similarity clustering, etc., to perform topic merging and trend statistics on texts. These methods can output a macro-level outline of "hot topics," but they lack interpretable modeling of the causal relationships, sequential relationships, and derivative branches between events, and cannot stably depict the chain structure of "core event - derivative event - re-propagation node."

[0019] At the commercial application level, some vendors' multi-source public opinion monitoring systems have achieved coverage of mainstream platforms and have undergone engineering enhancements in areas such as data collection frequency, media content transcription, and dialect processing, thereby improving the monitoring scope and visibility. However, from an algorithmic perspective, these systems still rely heavily on statistical features, dictionary matching, or fixed rules in event scoring and early warning decisions. They lack a learnable update mechanism for the changing importance of different key features at different stages, and they also lack the ability to systematically incorporate the actual evolutionary feedback after the early warning into the closed-loop model update. Therefore, when facing complex public opinion characterized by high noise, multimodal mixing, and cross-platform diffusion, there is still room for improvement in the stability and transferability of early warnings.

[0020] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0021] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0022] Figure 1A flowchart illustrating an example of a tourism public opinion event identification method based on dynamic weight adjustment according to an embodiment of this application is shown.

[0023] Regarding the execution entity of the method in this application, it can be any controller or processor with computing or processing capabilities, such as the controller of a tourism public opinion monitoring platform. A dynamic weight adjustment mechanism, constrained by real-world evolutionary feedback and driven by contribution, is constructed and embedded into a unified structured feature expression and weighted scoring framework, forming a continuously iterative event identification closed loop. Through this closed-loop mechanism, the system's judgment criteria are no longer static but automatically calibrated and migrated with the monitoring cycle, thereby maintaining more stable and reliable identification and early warning output under different public opinion forms and evolutionary stages.

[0024] In some examples, it may be integrated into an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device may be diverse.

[0025] like Figure 1 As shown, in step S110, multi-source tourism public opinion data associated with the target tourist destination is obtained.

[0026] Here, multi-source tourism public opinion data includes at least public opinion content data as well as release time information, data source identification information, publisher information and geographical location information associated with the public opinion content data. Public opinion content data includes text content and / or media content.

[0027] In some implementations, the system establishes collection channels to multiple public data sources, which may include social media platforms, short video / text communities, news websites, review and travel platforms, forums, and Q&A forums. Regarding the association of target tourist destinations, the system can employ a combined strategy of destination dictionaries, entity recognition, geofencing, and POI mapping. For example, it pre-maintains destination aliases, attraction / hotel / transportation hub names, landmarks, and common spelling variations, and performs named entity recognition on the collected content to locate destination entities; simultaneously, it spatially matches content with geographic tags with destination boundaries (fences) based on latitude and longitude; and it directly maps review / travel platform data to a destination list using unique POI identifiers. Furthermore, to improve recall, the system can introduce synonym expansion, hot word association, and contextual constraints (such as tourism scenario terms like "scenic spot / tickets / queues / road closures") to reduce missed or incorrect data collection based solely on keywords.

[0028] In addition to text content and / or media content, each piece of public opinion data collected also includes the associated publication time, data source identifier, publisher information, and geographic location information, and the fields are standardized when entering the database.

[0029] For example, time fields from different platforms are uniformly converted to the same time zone and timestamp format; data sources are solidified into "source IDs" in the form of enumeration / encoding; publisher information is standardized into computable fields (such as account type, authentication status, follower count range, historical posting frequency, etc.); and geographical location is unified into latitude and longitude or administrative division codes. This enables aligned analysis under a unified time, space, and platform identification system, providing a consistent data foundation for cross-platform integration, propagation chain tracking, and risk assessment.

[0030] In step S120, the multi-source tourism public opinion data is cleaned and structured, and the public opinion content data is mapped into a structured feature representation with a unified dimension to generate a set of public opinion records.

[0031] Here, each public opinion record in the public opinion record set contains a structured feature representation as well as corresponding publication time information, data source identification information, publisher information, and geographical location information.

[0032] In some implementations, the system performs layered cleaning on the raw multi-source data. Specifically, the first layer is rule-based cleaning, including deduplication (same link / same text hash / nearly duplicate text), removal of abnormal characters and advertising templates, filtering of irrelevant topics (content containing only destination terms but semantically unrelated), and time / location anomaly verification (obviously unreasonable future times, latitude and longitude out-of-bounds, etc.). The second layer is quality control cleaning, including identification of suspected bot-generated posts (high frequency in a short period, extremely high content similarity, abnormal account profiles), identification of copied and mirrored content (content repeatedly spread across multiple platforms but without new information is marked as "quoted content"), and adding noise labels to records that cannot be completely removed but may affect the judgment for subsequent feature use. For media content, the system can extract usable structured clues, such as titles, tags, and platform-generated descriptive text for text and images; titles, descriptions, and topic tags for videos; and shooting time and platform-side descriptive information for images. The above processing does not require changing the media itself, but rather structuring the computable semantic clues carried by the media as much as possible.

[0033] After cleaning, the system maps public opinion content into a structured feature representation with a unified dimension. Optionally, text content can undergo word segmentation / syntactic processing to generate semantic vectors (e.g., fixed-dimensional embeddings based on pre-trained language models), while outputting topic distribution, keyword weights, etc.; media content can generate cross-modal vectors (e.g., content vectors of images / videos or vectors obtained by fusing titles, tags, and transcribed text), and align them with text vectors to a unified dimensional space. Ultimately, each public opinion record forms a combined record of "structured content representation + contextual information such as time / source / publisher / location." This transforms heterogeneous inputs into a homogeneous, measurable, and aggregable representation, significantly reducing the incomparability caused by cross-platform and cross-modal differences.

[0034] In step S130, multi-dimensional key evaluation features for event identification are extracted from each public opinion record in the public opinion record set based on the preset feature definition, so as to generate the corresponding key evaluation feature vector set.

[0035] In some implementations, the system extracts multi-dimensional key evaluation features for event identification from each public opinion record based on preset feature definitions. These features cover dimensions such as emotion, dissemination impact, credibility, spatiality, and novelty, in order to support comprehensive judgment of emergencies in tourism scenarios.

[0036] For example, the emotion dimension can include emotional polarity (positive / negative / neutral), emotional intensity (such as intensity scores for anger, fear, and complaints), and the density of negative trigger words; the dissemination dimension can include the interaction growth rate per unit time, the depth of the forwarding / comment chain, and the frequency of recurrence across platforms; the influence dimension can include the publisher's influence (range of followers, certification type, and historical content dissemination level), and the level at which the content is cited; the credibility dimension can include the source credibility level (official media / certified accounts / ordinary users), content consistency (the degree of consistency in the narratives of the same event from different sources), and abnormal dissemination patterns (suspicious characteristics of online trolls); the spatial dimension can include the distance from the core area of ​​the destination and the density of postings in the same geographical area; novelty can include the degree of difference from the themes of previously known events and whether new risky keyword combinations have appeared. These features can be calculated directly from structured fields or further derived from content vectors, such as outputting the probability of event types like complaints / accidents / weather / traffic / service quality through a classifier.

[0037] Furthermore, the various features can be scaled and missing data handled. For example, continuous values ​​can be normalized / segmented, categorical features can be one-hot or target-based encoded, and missing fields can be introduced with missing indicator bits (to distinguish between "true zero" and "not obtained"). The resulting set of key evaluation feature vectors corresponds one-to-one with a single public opinion record. This compresses the difficult-to-compare public opinion content and context into a computable and interpretable set of multi-dimensional indicators, reducing scoring bias caused by inconsistent feature dimensions.

[0038] In step S140, the feature weight vector corresponding to the current monitoring period is obtained, and the key evaluation feature vector set is weighted and combined based on the feature weight vector to obtain the event score result of each public opinion record; the event score result is compared with the event judgment threshold to identify potential tourism public opinion events, and if a potential tourism public opinion event is identified, an early warning message is generated and the event identifier associated with the early warning message is recorded.

[0039] In some implementations, the system runs an event recognition process according to a monitoring cycle (e.g., 5 minutes, 15 minutes, or 1 hour), and obtains the corresponding feature weight vector at the beginning of each cycle. The weight vector may be derived from the system initialization configuration (cold start) or the update result of the previous cycle, and upper and lower limits and smoothing factors can be set to avoid drastic weight oscillations.

[0040] For each public opinion record, the event score can be obtained based on a weighted combination of "weight vector × key assessment feature vector". For example, a unified risk score can be formed by linear weighted summation or non-linear compression (such as Sigmoid). In addition, a time decay factor can be introduced to certain features before weighting, so that content closer to the current moment contributes more to the score, thus better meeting the needs of real-time monitoring.

[0041] To avoid misjudgments triggered by outliers in a single piece of content, further aggregation scoring can be performed on content with the same theme or high similarity within a period. For example, candidate event clusters can be formed first by semantic similarity, time window, and geographical proximity, and then "event-level scores" can be formed by statistical measures of the scores within the cluster (mean / quantile / peak duration). Public opinion records can be bound to event clusters to form event identifiers.

[0042] Subsequently, the system compares event scores with event judgment thresholds to identify potential tourism-related public opinion events. When trigger conditions are met, an early warning message is generated, and the event identifier associated with the early warning message is recorded (for subsequent closed-loop tracking). The early warning message may include an event summary (keywords / type probability), key evidence content (high-scoring post samples), dissemination trend (growth curves over the past N periods), spatial distribution (heat or administrative region clustering), and suggested points of concern (e.g., "suspected traffic congestion / service complaints / safety accidents"). Thus, under a unified scoring framework, multi-dimensional clues are transformed into thresholdable risk signals, and subsequent data feedback and model calibration are bound to the same event object through event identifiers, ensuring that early warning outputs are traceable and can be continuously tracked.

[0043] In step S150, after the warning, the actual event evolution feedback data corresponding to the event identifier is collected, and the identification accuracy is calculated based on the deviation between the actual event evolution feedback data and the warning information.

[0044] Here, after the warning is triggered, the system starts the event tracking and collection mechanism, continuously gathering relevant public opinion and external feedback information over a period of time around the event identifier, forming actual event evolution feedback data.

[0045] In some implementations, the feedback data may include not only the subsequent spread trajectory of public opinion on the same topic (changes in popularity, expansion or contraction of the spread, changes in the main information sources), but also credible anchor information related to the event, such as announcements from the destination management, reports from authoritative media, and platform-side handling results (debunking marks, content removal).

[0046] When calculating the recognition accuracy, the system aligns and compares key predicted quantities in the early warning information with actual evolutionary feedback. Optionally, the accuracy can be divided into "event existence accuracy" (whether the warned event is ultimately confirmed / continues to escalate), "severity bias" (the difference between the predicted risk score and the actual impact), and "time sensitivity bias" (the lead time of the early warning point relative to the event peak or confirmation point), and can be merged into a single recognition accuracy index according to a preset method. Thus, the reliability of early warnings is quantified based on evolutionary facts, avoiding the need for weight adjustments to rely solely on experience or static rules.

[0047] In step S160, the contribution of each key evaluation feature is calculated based on the recognition accuracy, and the feature weight vector is dynamically updated according to the contribution to obtain the updated feature weight vector for the next monitoring period; wherein, the key evaluation feature with a greater contribution is assigned a higher weight in the updated feature weight vector.

[0048] In some implementations, the system calculates the contribution of each key evaluation feature based on the recognition accuracy. This contribution characterizes the marginal effect of a feature on correct recognition in the current period.

[0049] The contribution can be achieved in several equivalent ways. For example, each feature can be perturbed / ablated (set to zero or replaced with the mean) and the decrease in accuracy can be observed, with the decrease as the contribution of the feature; or the correlation strength between the feature and the "correct / incorrect warning" can be calculated on the event-level aggregate score, and the contribution can be formed by combining the stability of the feature in high-scoring samples; or an approximate interpretable method (such as local linear interpretation) can be used to obtain the feature importance score.

[0050] To avoid overestimating noise features during random periods, the system can introduce confidence corrections to the contribution, such as combining sample size, event duration, and information source credibility distribution to attenuate or weight the contribution, thereby making the contribution more robust.

[0051] Preferably, when updating the weight vector, the system can dynamically allocate weights based on their contribution, and normalize the weights to ensure the sum is 1 or falls within a controllable range. Simultaneously, a smooth update strategy can be introduced to suppress periodic fluctuations, and upper and lower limits can be set to prevent a feature weight from being pushed to near zero, causing it to lose its ability to detect subsequent changes. The updated weight vector serves as the weight input for the next monitoring cycle. The scoring mechanism forms a closed-loop iteration between cycles, ensuring that features with greater contributions have a higher proportion after the update, thus dominating the judgment in subsequent scoring.

[0052] Therefore, the scoring criteria can be automatically calibrated as the information environment and event stage change, achieving adaptive migration while maintaining a unified scoring framework. This improves the identification stability and early warning reliability under continuous monitoring conditions, and reduces the risk of stage mismatch caused by fixed weights.

[0053] Regarding the processing details of structured feature representation, in some examples of embodiments of this application, statistically based abnormal information source filtering is performed on the publisher information, the publishing frequency index and content duplication rate index within a preset sliding time window are calculated, and publishers whose indexes exceed a preset abnormal threshold are marked as noise sources, and public opinion data published by noise sources within the preset sliding time window are removed.

[0054] Here, to avoid systemic disturbances to event ratings and subsequent weight updates caused by frequently spamming accounts, marketing accounts, or reposting accounts, statistically based abnormal information source filtering is performed on the publisher information.

[0055] Specifically, using the publisher's identifier as the granularity, the system calculates the frequency of publication (e.g., the number of public opinion records per unit time) and the content duplication rate within a preset sliding time window (e.g., 10 minutes, 30 minutes, or 1 hour, configurable according to monitoring granularity). The content duplication rate is obtained by aggregating the similarity of the publisher's text / media summary fingerprints within the window. For example, it calculates hash fingerprints or semantic vector similarity for text, and perceptual hashes for media, and counts the proportion of highly similar samples, thus obtaining a quantitative characterization of "repeated publication." When the publication frequency and / or content duplication rate exceed a preset abnormal threshold, the publisher is marked as a noise source, and its public opinion data published within the sliding window is removed (the removal mark is retained for auditing if necessary). This mechanism reduces the bias caused by extreme interaction volumes and duplicate content stacking from abnormal sources at the structured mapping front end, making subsequent feature vectors more reflective of the true distribution of public opinion subjects and reducing the risk of misjudgment and score drift triggered by noise.

[0056] Then, the filtered text content is preprocessed by word segmentation and stop word removal, and the dialect correction model is used to identify non-standard dialect words in the text. Based on semantic similarity, the non-standard dialect words are mapped to standard semantic words. Then, the pre-trained language model is called to encode the processed text to generate text embedding vectors.

[0057] Specifically, the filtered text content undergoes word segmentation and stop word removal preprocessing to remove redundant components such as function words and filler words that contribute little to event recognition. Emojis, special characters, and homographs are then replaced to improve the consistency of text encoding. Subsequently, a dialect correction model is invoked to identify non-standard dialect words in the text. This model can output the positions and confidence scores of suspected dialect words using sequence labeling or lexical classification, and maps non-standard dialect words to standard semantic words at the lexical level based on semantic similarity (e.g., selecting the candidate words with the highest similarity from a standard vocabulary or synonym set), thereby reducing feature sparsity caused by differences in the expression of the same meaning in different regions. The corrected text is then fed into a pre-trained language model for encoding, generating text embedding vectors (such as sentence vectors or segment vectors), making text features measurable and comparable at the semantic level. Thus, without relying on manual rules, the standardization and semantic alignment of text expression are improved, thereby enhancing the identifiability and robustness of event-related text signals.

[0058] Next, multi-channel feature extraction is performed on the media content, including: calling an optical character recognition algorithm to extract scene text from image or video frames, and converting the scene text into auxiliary text vectors; calling a visual analysis model to extract visual content features from image or video frames to generate visual embedding vectors; and calling a speech emotion recognition model to extract acoustic emotion features from the video audio stream.

[0059] Specifically, for image or video frames, optical character recognition algorithms are used to extract scene text (such as road signs, announcements, ticketing information, and on-site slogans). The identified scene text is then converted into auxiliary text vectors using an encoding process consistent with the text content to strengthen the support of explicit textual evidence in the image for event identification. For visual content, a visual analysis model is used to extract features from image or video frames, obtaining visual embedding vectors that can represent scene elements, crowding levels, and abnormal phenomena (such as queues, closures, and accident scenes). Video scenes can obtain stable representations through keyframe sampling or temporal pooling. For videos with audio tracks, a speech emotion recognition model is used to extract acoustic emotion features from the audio stream (such as emotion intensity / polarity correlation vectors obtained based on speech rate, energy, pitch, and spectral features). When a certain channel of the media is missing (e.g., no audio track), zero vectors or masking mechanisms can be used to maintain dimensional consistency. Thus, readable textual evidence, visible scene evidence, and audible emotion evidence in media content are structured in parallel, enabling information related to the event but not explicitly expressed in the main text to enter a unified scoring chain, improving the coverage and sufficiency of evidence for complex public opinion materials.

[0060] A multimodal feature fusion space is constructed, and text embedding vectors, auxiliary text vectors, visual embedding vectors, and acoustic sentiment features are projected onto a shared semantic dimension space using their respective learnable projection matrices.

[0061] Specifically, to enable the computability of different modal vectors at the same semantic scale, a multimodal feature fusion space is constructed, and dimensional alignment is achieved through their respective learnable projection matrices. Specifically, let the original dimensions of the text embedding vector, auxiliary text vector, visual embedding vector, and acoustic sentiment feature be respectively... Construct a learnable projection matrix , , , Project each modality vector onto a shared semantic dimension. The aligned modal representations are obtained. If necessary, bias terms can be introduced and nonlinear transformations can be performed to enhance expressive power. This projection matrix can be jointly optimized during offline training using objective functions related to event recognition, thereby enabling each modality to form an aligned semantic coordinate system in a shared space. This avoids the "dominance of a particular modality" or "scale imbalance" caused by directly splicing original modal vectors with different dimensions and distributions, allowing subsequent weighted scoring and contribution attribution to operate stably on the same semantic dimension benchmark.

[0062] Normalization is performed on each modality vector after projection, and feature concatenation is performed to generate a structured feature representation of the corresponding public opinion record. .

[0063] Specifically, normalization can be achieved by... Normalization or layer normalization is performed to constrain the amplitude range of the alignment vectors of each modality and reduce the influence of outliers. Subsequently, the normalized multimodal vectors are concatenated in a preset order to form a structured feature representation with a unified dimension. This information, along with the record's publication time, data source identifier, publisher information, and geographic location information, is stored in the public opinion record set. This ensures that different records are comparable and reusable within the same representation space.

[0064] Regarding the implementation details of extracting multidimensional key evaluation features, in some examples of embodiments of this application, the focus is on the first... Each public opinion record is used to obtain its corresponding structured feature representation. And associated attribute information.

[0065] In some implementations, the associated attribute information includes at least the publication timestamp, publisher attributes (such as number of followers, authentication type, historical interaction level, etc.), public opinion interaction statistics (such as the number of reposts and comments within a preset observation window), and geographic location information. By binding the semantic expression of public opinion content with its dissemination behavior, source, and spatiotemporal attributes, various key evaluation features can be calculated based on a unified data foundation of the same public opinion record. This achieves integrated "semantic-dissemination-source-spatiotemporal" modeling input for a single public opinion record, improving the repeatability and cross-data source consistency of key evaluation feature calculations, and reducing computational noise caused by missing fields or inconsistent formats.

[0066] Representing structured features Input a pre-set sentiment analysis model for semantic decoding to calculate the sentiment intensity of public opinion records. And calculate the influence based on the associated attribute information. ,growth rate Source credibility and spatiotemporal proximity In order to construct the corresponding multidimensional key evaluation feature vector.

[0067] In some examples, structured features are represented Input a pre-defined sentiment analysis model for semantic decoding to output the sentiment intensity of the public opinion record. The sentiment analysis model can be a pre-trained language model or a multimodal sentiment model, and its input is... or The text semantic subvectors are used as the output, which are the probability and intensity scores of sentiment polarity. In practice, the negative sentiment probability, strong dissatisfaction probability, or sentiment intensity regression value output by the model can be used as... The basis for calculation, thus It can continuously represent the "intensity of emotion" rather than just providing positive or negative labels.

[0068] Among them, influence The calculation is performed by constructing a weighted logarithmic interaction model, which uses a logarithmic function to smooth the publisher attribute data to suppress long-tailed distribution differences. Equation (1) In the formula, The number of followers listed in the publisher's information is used to represent the publisher's basic influence. The number of reposts of public opinion content within a preset observation window is used to characterize the breadth of dissemination; The number of comments on public opinion content within a preset observation window is used to characterize the depth of interaction; and These are the breadth of dissemination weighting coefficient and the depth of interaction weighting coefficient, respectively, used to balance the contributions of "forward-driven diffusion" and "comment-driven discussion".

[0069] In equation (1), the logarithmic function is used because the public opinion dissemination index has a significant long-tail distribution. Directly using the raw count will lead to top accounts or extreme viral samples dominating the score, while log smoothing can compress the scale of extreme values, so that medium-sized but continuously spreading events can still be effectively identified; in addition, through and The introduction of this feature can be adjusted in a controllable manner based on business priorities (such as a greater focus on dissemination and diffusion or a greater focus on interactive disputes).

[0070] Therefore, without losing strong propagation signals, the scoring imbalance caused by long-tail distribution is suppressed, so that the influence characteristics can reflect the sudden spread at the top and retain the risk warning ability of the middle and lower tiers, thereby improving the comparability of different platforms and account sizes.

[0071] Among them, spatiotemporal proximity The calculation is performed by constructing a spatiotemporal Gaussian coupled decay model to quantify the degree to which public opinion records converge toward the core event in the spatiotemporal dimension. Equation (2) In the formula, For geographic location information, Core coordinates of the target tourist destination (e.g., the center point of the scenic area or a geographical reference point defined by the management). This is a geodesic distance calculation function used to calculate... and Spatial distance between them; To publish time information, This refers to the sampling time of the current monitoring period or the end time of the current time window; This characterizes the decay of the timeliness of public opinion information. and These are the spatial dimension decay factor and the time dimension decay factor, respectively, used to adjust the decay rate of proximity due to distance and time difference.

[0072] In equation (2), an exponential decay mechanism is adopted. When the location of the public opinion release is far from the destination or the release time is far from the trigger time, its correlation with the core event should usually decrease rapidly to avoid interference from irrelevant cross-regional discussions and historical news on the identification of the current event; while when the public opinion is close to the destination in space and close to the trigger time in time, This will approach 1, thereby strengthening the role of such near-time and near-location samples in event identification. This will enable automatic suppression of "irrelevant public opinion from other locations" and "historical backflow public opinion," enhancing the system's ability to focus on real-world events and improving the accuracy of potential event location and the reliability of early warnings.

[0073] More specifically, in some examples of embodiments of this application, regarding the growth rate The calculations include: Using the exponential moving average model for the first The number of mentions of the public opinion information corresponding to each record at the current sampling time First-order difference rate Perform smoothing calculations: Using the exponential moving average model for the first The number of mentions of the sentiment topic or destination entity associated with each sentiment record at the current sampling time. First-order difference rate Perform smoothing calculations: Equation (3) In the formula, For the public opinion topic or destination entity at the previous sampling time smooth growth rate This is the decay smoothing coefficient. The system will calculate the trend growth rate. This public opinion record is assigned as a characteristic of its growth rate.

[0074] Regarding the growth rate In this embodiment, instead of simply using instantaneous counting of public opinion, an exponential moving average (EMA) model was constructed based on formula (3) to handle the first-order difference rate of public opinion mentions. .

[0075] More specifically, raw public opinion data is often accompanied by random noise (such as crawler jitter or instantaneous spikes caused by online trolls), and directly using the instantaneous rate can easily lead to false alarms in the system. Formula (3) introduces an attenuation smoothing coefficient. A "memory mechanism" was established, in which The value depends not only on the burst speed at the current moment. It also retains the trend from the previous moment. Inertia. When When the setting is larger, the model focuses more on the stability of long-term trends; when When the settings are smaller, the model is more sensitive to sudden changes. This enables smoothing filtering of the public opinion heat curve, which can effectively suppress high-frequency noise interference and keenly capture the real acceleration of the transition of public opinion events from the incubation period to the outbreak period.

[0076] Regarding emotional intensity The calculations include: Construct an attention-based semantic decoding model for structured feature representation. A joint analysis of emotional polarity and intensity: Equation (4) In the formula, This is a self-attention mechanism function used to capture long-distance semantic dependencies within feature vectors; and These are the weight matrix and bias term of the semantic decoding model, respectively; It is a hyperbolic tangent activation function, whose absolute output value represents the intensity of emotion from calm to agitation.

[0077] It should be noted that the input structured feature representation It is often a high-dimensional vector containing long text sequences or multimodal information, where the sentiment is often determined by a few key features (such as specific negative adjectives or intense speech segments).

[0078] Therefore, a self-attention mechanism function is introduced into equation (4). This dynamically allocates weights, automatically focusing on key information fragments within the feature vector that have long-distance semantic dependencies. Subsequently, a linear transformation maps the focused features to the sentiment semantic space, and utilizes... The function compresses the output to the interval [-1, 1] to represent polarity (-1 for extreme negativity, 1 for extreme positive). Finally, it performs an absolute value operation. The intensity of emotions. (From 0 to 1). It should be noted that the direction (polarity) of the emotion is determined by the sign before the absolute value operation, while... It focuses on accurately quantifying the "intensity" of emotions, enabling the system to distinguish between "ordinary complaints" and "extreme anger," and ensuring that a high-level warning can be quickly triggered through a dynamic weighting mechanism when high-intensity negative emotions are detected.

[0079] Regarding source credibility In some implementations, the calculation can employ a multi-factor weighted evaluation strategy. Specifically, the system first determines the platform type weight based on the data source identification information. (For example, official news media are given high weight, while anonymous forums are given low weight); secondly, the account verification level is analyzed based on the publisher information. (For example, verified users with a "blue V" badge are considered more trustworthy than unverified users); finally, an authenticity score is calculated based on the publisher's historical data. (For example, accounts that have a history of spreading rumors will be demoted.)

[0080] The system uses linear combination Calculate the final score, where The confidence coefficients for each factor are used to construct an "information filter." In the later stages of a public opinion crisis, when rumors proliferate or online trolls flood the internet, the dynamic weighting algorithm automatically increases the source credibility. The weighting ratio is used to effectively suppress the interference of low-confidence noise on the qualitative analysis of events, ensuring that early warning decisions are based on reliable information sources.

[0081] Regarding the implementation details of calculating and updating the feature weight vector, in some examples of embodiments of this application, the real event labels are determined based on actual event evolution feedback data, and a differentiable cross-entropy loss function is constructed in combination with the event scoring results to quantify the difference between the predicted distribution and the real distribution within the current monitoring period.

[0082] Then, sensitivity analysis is performed on the feature weight vector by calculating the cross-entropy loss function with respect to the first... The partial derivatives of the weights of each key evaluation feature are used to determine the gradient contribution of that feature within the current monitoring period. The gradient contribution value represents the marginal utility of the feature weight changes in reducing prediction error.

[0083] Here, in calculating the gradient contribution value In its implementation, instead of simple error backpropagation, an evaluation system based on sensitivity analysis was constructed. Specifically, a differentiable mapping relationship was established between "early warning deviation" and "feature weights." The system uses the collected actual evolutionary feedback (such as the post-confirmed event truth level) as the ground truth label and constructs a differentiable cross-entropy loss function to quantify the difference between the predicted distribution and the actual distribution within the current monitoring period. Subsequently, the difference is calculated using this loss function with respect to the [missing information]. Feature weights The partial derivatives are used to obtain the gradient contribution value. .

[0084] If the weight of this feature is fine-tuned, how much will the rate of decrease in the overall recognition error of the system (i.e., marginal utility) be? This will transform qualitative business feedback (such as "this was a false alarm") into quantitative mathematical gradients, enabling the system to accurately identify which feature (e.g., over-computation of sentiment or misjudgment of growth rate) caused the current recognition bias, thus providing precise guidance for the targeted correction of the weights.

[0085] A multidimensional attention softmax mechanism based on exponential weighting is used to update the feature weight vector through nonlinear mapping, so as to achieve adaptive weight allocation based on feature importance: Equation (5) Among them, the normalized partition function Defined as: Equation (6) In the formula, and These represent the current monitoring period. The Middle The first feature and the first The weight values ​​of each feature, Indicates the next monitoring cycle The Middle Update weights for each feature. The learning rate parameter is used to control the step size of the weight adjustment. A set of type identifiers for multidimensional key evaluation features. These represent emotional intensity, influence, growth rate, source credibility, and spatiotemporal proximity, respectively. Represents a set Any feature type index in the, Indicates the first The gradient contribution value corresponding to each feature type.

[0086] For the dynamic updating of the feature weight vector, the multidimensional attention softmax mechanism based on exponential weighting in formulas (5) and (6) can be used. In formula (5), the exponential function is used. As a nonlinear amplifier, the learning rate Control the aggressiveness of the adjustment; if the gradient contribution value of a certain feature... A larger value (indicating that the feature is very helpful in reducing error) will result in its weight being amplified exponentially, while a smaller value will be suppressed rapidly. Meanwhile, the normalized partition function defined by formula (6) Sum and normalize all updated components to ensure that the new weight vector always satisfies the probability distribution constraint (i.e., the sum is 1).

[0087] This approach simulates the attention allocation mechanism in deep learning, enabling the system to adaptively prioritize features without human intervention. Specifically, it automatically focuses on growth rate and sentiment intensity in the early stages of a public opinion crisis, while automatically shifting weights to "source credibility" during the fermentation phase of rumors. Thus, through a non-linear dynamic redistribution mechanism, the model's adaptability to the complex and ever-changing lifecycle of public opinion is greatly enhanced, effectively solving the problem of false negatives or negatives caused by the unchanging nature of traditional static weight models.

[0088] In some examples of embodiments of this application, after the Softmax nonlinear mapping, adaptive time momentum smoothing is performed on the updated feature weight vector to suppress parameter oscillations.

[0089] More specifically, establish a length of The sliding observation window is used to calculate the first... The temporal variance of gradient contribution values ​​within each monitoring period is used to construct a gradient instability index reflecting the system's evolutionary state. : Equation (7) In the formula, Features The mean within the sliding window, The total dimension of the features. Indicates the first The first monitoring cycle Gradient contribution value of each feature; gradient instability index Used to quantify the degree of fluctuation in the feature contribution on a periodic sequence.

[0090] Here, in the specific implementation of suppressing parameter oscillations, the gradient instability index is first constructed using formula (7). The system established a length of A sliding observation window (e.g., 5 periods or 10 periods) is used to store the most recent observations. The gradient contribution values ​​of each feature within each monitoring period are calculated. These gradient values ​​are then compared to the mean. The temporal variance is calculated and averaged over all feature dimensions.

[0091] Specifically, the gradient direction at a single moment is highly susceptible to interference from data noise (such as instantaneous data distortion caused by online manipulation of rankings), while variance can capture this high-frequency "reciprocating jitter." If A smaller value indicates that the feature importance evaluation has been stable in the recent past, and the gradient direction is consistent; if A significant increase indicates that the system's evaluation of the features is fluctuating repeatedly (i.e., the gradient fluctuates frequently and significantly or changes sign). This provides the system with a quantified state entropy value, enabling it to accurately distinguish between "real changes in public opinion trends" and "instantaneous random noise disturbances," preventing the algorithm from falling into a chaotic state due to oversensitivity to noise.

[0092] Then, a nonlinear momentum adjustment function is constructed based on the gradient instability index. Dynamic calculation of the first Momentum smoothing coefficient for each monitoring period : Equation (8) In the formula, and These are the lower and upper bounds of the momentum smoothing coefficient, respectively. The preset stability threshold, The gain coefficient is used to adjust the response sensitivity. The adjustment function ensures that the system remains sensitive to high fluctuations. Approaching To enhance the inertia of historical weighting.

[0093] Here, the hyperbolic tangent function is used. Combination The function constructs a S Type-saturated response curve. By setting a stability threshold. When the system is unstable When it is below this threshold, The function outputs 0, making Maintain at the lower bound (For example, 0.1), at this point the model is in "agile mode," able to quickly adapt to new weight changes; once the instability exceeds the threshold, The function value rises rapidly and approaches 1, forcing Smoothly approaching the upper bound (For example, 0.9) switches the model to "conservative mode." Through variable structure control, an adaptive balance between system robustness and sensitivity is achieved. Thus, during periods of stable public opinion, low momentum ensures that the model can keenly capture weak crisis signals; while during periods of violent fluctuations in public opinion (such as when intense confrontation between positive and negative evaluations leads to gradient chaos), high momentum forces the system to maintain "inertia," avoiding avalanche-like abrupt changes in weight parameters due to individual extreme samples, thereby ensuring the long-term operational stability of the early warning system.

[0094] Furthermore, based on the momentum smoothing coefficient Perform weighted smooth updates: Equation (9) In the formula, For the next monitoring cycle The Middle The final effective smoothing weights of each feature, The next monitoring period is obtained by using an exponentially weighted multidimensional attention softmax mapping. The Middle Update weights for each feature. For the current monitoring period The Middle Historical smoothing weights for each feature.

[0095] The final weighted smoothing update is performed through recursive filtering using formula (9), which maps the current period's "suggested update weights" obtained from the Softmax mapping. Compared to the "historical smoothing weight" of the previous period Perform linear weighted fusion.

[0096] The key here is that the weighting coefficient is dynamically calculated by equation (8). When the system determines that the current environment is unstable ( When the value is large, equation (9) will retain more of the original value. (Historical experience) suggests avoiding drastic parameter jumps; conversely, adopting more flexible approaches. (New Knowledge). The technical effect of this step is to add a "safety valve" to the dynamic weight adjustment mechanism, ensuring that the feature weight vector output to the next monitoring cycle is safe. It incorporates the latest feedback information and has sufficient temporal continuity, thereby generating a smooth and interpretable weight evolution trajectory and avoiding false alarms caused by weight jumps.

[0097] In some examples of embodiments of this application, a dynamic semantic event chain graph can also be constructed using smoothed feature weight vectors processed by adaptive time momentum smoothing to track the evolution path of public opinion.

[0098] In constructing a dynamic semantic event chain graph, the first step is to perform clustering operations on event nodes. For example, potential tourism public opinion data identified within the current and historical monitoring periods are input into a clustering algorithm (such as density-based DBSCAN or hierarchical clustering). This aggregates semantically similar and spatiotemporally adjacent fragmented public opinion records into discrete "event nodes." Each event node represents a key state in the evolution of public opinion (e.g., "tourist complaint stage," "official intervention stage," or "emotional outburst stage"). This transforms unstructured, massive data streams into structured graph nodes, providing the basic computational units for subsequent evolutionary path analysis. It achieves granular regularization of public opinion information, enabling the system to transcend single time slices and track the morphological changes of the same event at different lifecycle stages from a macroscopic perspective.

[0099] More specifically, potential tourism-related public opinion events are clustered into a set of event nodes, and for any two event nodes... and , using the Smoothing weights for each monitoring period Calculate event nodes and Cross-feature weighted cosine similarity between .

[0100] Equation (10) In the formula, and Representing event nodes respectively and In the Aggregated feature values ​​on class features, A set of type identifiers for multidimensional key evaluation features. For the next monitoring cycle The Middle The final effective weights for each feature are smoothed; the similarity formula is applied through weights. The feature space is dynamically stretched so that features with high weight in the current period play a dominant role in similarity measurement.

[0101] For calculating the correlation strength between event nodes, a cross-feature weighted cosine similarity algorithm as shown in Equation (10) was adopted. Traditional cosine similarity treats all feature dimensions equally, while in Equation (10), the weights after adaptive smoothing are used. The feature space was nonlinearly modified. Specifically, the numerator of the formula calculates the weighted inner product, while the denominator calculates the weighted Euclidean norm (i.e., the weighted modulus).

[0102] Therefore, a "dynamic stretching of the feature space" is adopted, that is, if the system determines that "emotional intensity" and "growth rate" are the dominant features (i.e., weights) in the current period, then... If the similarity is relatively high, then Equation (10) will amplify the impact of the numerical differences between these two dimensions on the final result when calculating similarity, while suppressing the interference of low-weight features such as "source credibility". Thus, the criteria for determining the relevance of events are no longer static and rigid, but can be adaptively adjusted with changes in the public opinion environment. That is, during the period dominated by emotions, nodes with similar emotions are more likely to be connected; during the period dominated by dissemination, nodes with similar dissemination patterns are more likely to be connected, thereby ensuring that the generated event chain conforms to the current public opinion evolution logic.

[0103] Then, based on cross-feature weighted cosine similarity Construct a directed acyclic graph and assign similarity Event nodes that exceed a preset link threshold are connected sequentially to generate an event chain graph.

[0104] Here, a directed acyclic graph (DAG) is constructed based on the calculated similarity matrix. Specifically, the system traverses all event node pairs. If their similarity The number of links exceeds the preset link threshold, and the node... The occurrence time is earlier than the node If so, then a directed edge is established between the two.

[0105] By leveraging the similarity of high-dimensional features to infer causal evolutionary relationships, an event chain graph can be constructed that intuitively displays the origins and development of public opinion. At the application level, this can help regulatory authorities quickly identify the "evolutionary path" of complex public opinion events. For example, it can clearly reconstruct how a "tourist rip-off incident" evolved from "sporadic negative reviews (node ​​A)" to "influencer reposts (node ​​B)" and ultimately led to "nationwide boycotts (node ​​C)," providing intuitive and visual decision-making basis for tracing the source of public opinion and judging trends.

[0106] In some examples of embodiments of this application, after generating an event chain graph, key public opinion evolution paths can be extracted and key points of public opinion mutations can be identified based on graph topology analysis.

[0107] More specifically, define event nodes in the event chain graph. Cumulative path energy state A state transition equation is constructed using a dynamic programming algorithm to search for the globally optimal main public opinion evolution path. : Equation (11) In the formula, This represents all pointers to nodes in a directed acyclic graph. The set of predecessor nodes, express Any node in, For event nodes The cumulative path energy state, For nodes The influence characteristic value, Path energy adjustment factor; Represents event nodes and Cross-feature weighted cosine similarity between them.

[0108] Here, after generating the event chain graph, in order to sort out the main thread of event development from the complex network structure, a state transition equation based on dynamic programming is constructed using formula (11).

[0109] In equation (11), the problem of finding the main path is transformed into an optimization problem of finding the path with the maximum cumulative energy. Specifically, the state It consists of two parts, one of which is the inherent properties of the node itself. (That is, the greater the influence of the event node, the more likely it is to be on the main path), and another part comes from its predecessor node. Maximum path energy transmitted Regulatory factors The weight used to balance "node popularity" and "evolutionary coherence".

[0110] Then, by backtracking... The sequence of nodes that reaches the global maximum value determines the main public opinion evolution path. , The first in the main public opinion evolution path 1 node The total number of nodes in the main public opinion evolution path.

[0111] In some implementations, the system traverses the nodes in the graph according to topological sorting, calculates and records the optimal predecessor pointer for each node, and finally uses a backtracking algorithm to trace back from the end point to the beginning point to obtain the node sequence. This enables intelligent noise reduction and connection of fragmented public opinion information, automatically ignoring minor details with low influence or weak relevance, and accurately reconstructing the core framework of a public opinion event from its "cause" to its "climax" and then to its "outcome," providing managers with a clear view of the event's retrospective.

[0112] Furthermore, along the main path of public opinion evolution Perform sentiment gradient detection and calculate the discrete sentiment gradient between adjacent nodes. : Equation (12) In the formula, and The first on the path The node and the first The emotional intensity feature value of each node, The time interval between events corresponding to the two nodes. The preset regularization parameter is used to avoid zero denominators and improve the numerical stability of gradient calculation; it will satisfy... nodes Identified as a key point of sudden change in public opinion, among which This is the preset mutation detection threshold.

[0113] In some implementations, this embodiment locates public opinion risk points along the evolutionary path. Perform sentiment gradient detection and use Equation (12) to calculate the discrete sentiment gradient between adjacent nodes. .

[0114] By simulating the process of differentiation, the molecule... The absolute change in emotional intensity between two adjacent event nodes was calculated, and the denominator was... This introduces the time dimension, in which This is a regularization term used to prevent division by zero errors caused by extremely short time intervals.

[0115] Therefore, we quantify the "acceleration" of emotional changes rather than simply their "amplitude." More specifically, we calculate... Compared with the preset mutation threshold The comparison is performed, and once the threshold is exceeded, the node is determined. This is the key point of mutation. Thus, it can keenly distinguish between "gradual dissatisfaction" and "explosive crisis"; even if the emotions of both nodes are negative, if the emotional intensity drops sharply from -0.2 to -0.9 in a short period of time, Equation (12) will generate a huge gradient value, thereby triggering a high-level alarm, so that the system can accurately capture the "turning point" in the process of public opinion fermentation (such as the reversal of public opinion caused by an official announcement), and help the regulatory authorities seize the best time to intervene.

[0116] In some examples of embodiments of this application, the event determination threshold can also be adaptively optimized based on the F1 score maximization criterion.

[0117] Specifically, utilizing the next monitoring cycle Smoothing weights Recalculate the event scores of the historical validation set samples and iterate through multiple candidate thresholds within the preset interval.

[0118] In some implementations, during the threshold adaptive optimization process, the system first utilizes the next monitoring period calculated in the previous stage. Smoothing weights The event scores are recalculated for samples in the historical validation set.

[0119] Since the weight vector has been updated (e.g., the weight of growth rate is increased, and the weight of source credibility is decreased), the scores of historical samples under the old weights can no longer represent their performance under the new model. By recalculating the new weights, it is ensured that the selection of the threshold is based on the latest feature importance distribution. Subsequently, the system generates a series of candidate thresholds within a preset probability interval (e.g., [0.3, 0.9]) with a fixed step size (e.g., 0.01), providing a search space for subsequent performance evaluation and ensuring that the system is always at the optimal decision benchmark.

[0120] For each candidate threshold, the corresponding number of true positives is counted. False positives and false negatives And calculate the accuracy. and recall rate : Equation (13) Based on accuracy and recall rate Calculate the harmonic mean index F1 score : Equation (14) Choose the option that makes the F1 score The candidate threshold that is maximized is set as the next monitoring period. The event judgment threshold is set to achieve a dynamic balance between false alarm rate and false negative rate.

[0121] For performance evaluation of candidate thresholds, the accuracy is calculated using equation (13). and recall rate The F1 score is calculated using equation (14). Equation (13) quantifies precision (accuracy) and recall (completeness).

[0122] Equation (14) uses the harmonic mean instead of the arithmetic mean. Specifically, the harmonic mean is very sensitive to minima. If a certain threshold results in extremely high precision but extremely low recall (i.e., missing a large number of real events to avoid false positives), or extremely high recall but extremely low precision (i.e., generating a large amount of noise to avoid false negatives), the F1 score calculated by Equation (14) will be lowered. Only when... and The F1 score reaches its maximum only when both are maintained at a high level and relatively balanced. This forces the optimization algorithm to find a compromise point, effectively preventing the system from falling into an extreme operating state.

[0123] Finally, this embodiment selects the candidate threshold that maximizes the F1 score and fixes it for the next monitoring cycle. The system establishes an event judgment threshold, thereby constructing a fully automated closed-loop feedback control system. When changes in the external public opinion environment lead to a decrease in the discriminative power of certain features, the F1 criterion automatically drives the threshold to move in a more robust direction. For example, during peak holiday travel periods, increased data noise may cause the system to automatically raise the threshold to maintain precision; while during sensitive periods, the system may slightly lower the threshold to prioritize recall.

[0124] This achieves a dynamic balance between false alarm rate and false negative rate, enabling the tourism public opinion early warning system to maintain its optimal performance without frequent manual intervention, significantly reducing operation and maintenance costs and enhancing the authority of the early warning.

[0125] Figure 2 A flowchart illustrating the operational mechanism of an example of a tourism public opinion event identification method based on dynamic weight adjustment according to an embodiment of this application is shown.

[0126] like Figure 2 As shown, the process first initiates a multi-source data perception workflow, collecting raw public opinion information from multiple dimensions such as social media posts, comments, image data, and geographic locations. This information is then cleaned and structured through a data preprocessing module. The core dynamic weight adjustment module, located at the center of the processing chain, adaptively allocates feature weights based on the current system state, thereby driving the feature extraction, event scoring, and detection modules to ensure accurate quantification and identification of potential public opinion events.

[0127] Furthermore, this system mechanism incorporates a self-evolving closed-loop feedback system. Event detection results not only generate alerts and reports as the final output, but are also fed back to the feedback weight update module to calculate the identification accuracy and optimize the weight parameters for the next cycle. Simultaneously, the regulatory constraint module imposes boundary restrictions on the weight update logic and detection results, ensuring that the system's operation conforms to both the evolution of public opinion and regulatory compliance requirements, thereby achieving fully automated collaboration from data perception and intelligent analysis to decision feedback.

[0128] To comprehensively verify the effectiveness and robustness of the proposed method, a multi-source heterogeneous tourism public opinion dataset covering Weibo, Douyin, Xiaohongshu, and professional tourism forums was constructed. This dataset contains complete public opinion records for multiple popular scenic spots over the past year, and the event categories and evolution stages were manually labeled by an expert group as a ground truth. The experiments used F1 score, accuracy, recall, and early warning response time as core evaluation metrics. In the comparative experiments, the following three schemes were set:

[0129] The Dynamic Weight method (DynamicWeight, the method in this application) fully adopts the dynamic weight adjustment and event chain construction algorithm proposed in this paper. Regarding parameter settings, the learning rate for feature weight updates is set to... The adjustment range of the adaptive momentum smoothing coefficient is set to Stability threshold The value was set to 0.05 to verify the model's adaptability under different public opinion fluctuations.

[0130] Static Weight Method: As a benchmark for ablation experiments, its feature weight vector is fixed throughout the entire cycle. (i.e., experience-weighted), without enabling gradient updates and momentum smoothing mechanisms, is used to verify the necessity of dynamic weight adjustments.

[0131] EventChain method: Constructs an event graph based on TF-IDF word frequency statistics and a fixed similarity threshold. It does not introduce multi-dimensional key evaluation features and dynamic weighting mechanism. It is used to compare the advantages of this application in complex semantic association mining.

[0132] Figure 3 The graph shows the comparison of F1 scores over consecutive evaluation periods according to different methods.

[0133] like Figure 3 As shown, it displays three performance evolution curves: the blue solid line with a circle indicates the dynamic weighting method of this application, the orange dashed line with a square indicates the static weighting method, and the green dotted line with a triangle indicates the event chain method.

[0134] Data comparison reveals that the dynamic weighting method proposed in this application exhibits significant performance advantages. Its F1 score starts at 0.74 and steadily increases with the evolution of public opinion, eventually converging stably in the high range of 0.79 to 0.81. This demonstrates that the adaptive weight update mechanism can effectively capture the phased characteristics of public opinion development. In contrast, the F1 score of the static weighting method mainly fluctuates between 0.63 and 0.66, lacking room for improvement; while the event chain method performs the weakest, hovering around 0.60 for a long time, making it difficult to cope with complex and ever-changing public opinion scenarios. The above comparative experimental results intuitively verify that this scheme, by introducing gradient contribution analysis and momentum smoothing mechanism, overcomes the limitations of fixed weight models and single word frequency models, thereby achieving higher event recognition accuracy and robustness.

[0135] Figure 4 The diagram shows the dynamic evolution curves of the weights of the five-dimensional key evaluation features according to embodiments of this application over a continuous evaluation period.

[0136] like Figure 4 As shown, the evolution trajectory of each feature weight exhibits significant stage-specific differences, accurately mapping the inherent laws of public opinion development. In the initial stage (cycles 1-5), the weight of influence (orange curve) remains at a high level of around 0.30, followed closely by emotional intensity (blue curve). This indicates that the system prioritizes key nodes with high dissemination power and negative emotional signals in the early stages of a public opinion outbreak to quickly pinpoint the core event source. As public opinion enters the diffusion and fermentation phase (cycles 10-17), the weight of source credibility (red curve) shows a significant upward trend, even surpassing other features to reach a peak of over 0.30. This trend verifies that the dynamic weight mechanism can adaptively suppress rumors and noise emerging in the middle stage, shifting the focus of judgment to the reliability of the information source. At the same time, the influence weight shows a "U-shaped" trend of first decreasing and then rebounding, reflecting the evolution of the dissemination subject from the initial core users to the general public, and finally back to authoritative summaries. In addition, spatiotemporal proximity (purple curve) shows a significant peak in the middle stage, reflecting the algorithm's dynamic attention to the convergence of the event's geographical location.

[0137] To thoroughly evaluate the independent contribution of each key assessment feature in event identification, we designed a feature mask ablation experiment. This involved forcibly setting the weights of specific features to zero and blocking their gradient update paths during dynamic monitoring. Experimental results show that the absence of source credibility and sentiment intensity had the most significant negative impact on system performance, causing a substantial decrease in F1 scores of 0.13 and 0.10, respectively. This data strongly demonstrates that "eliminating false sources" and "capturing intense negative emotions" are the core pillars for ensuring the accuracy of tourism public opinion identification. The removal of growth rate and influence caused an average decrease in F1 score of approximately 0.06, indicating that they played a crucial auxiliary role in quantifying the outbreak trend and dissemination level of events. The marginal utility of the absence of spatiotemporal proximity was relatively small, suggesting that it mainly provides fine-grained spatial positioning support under specific geographical constraints. The synergistic effect of these features together constitutes a complete view of public opinion identification.

[0138] This paper addresses the shortcomings of existing tourism public opinion monitoring technologies, such as insufficient multimodal fusion and recognition lag caused by static weight fixation. It proposes a public opinion event identification method based on dynamic weight adjustment. This method constructs a multidimensional key evaluation feature system encompassing sentiment intensity, influence, growth rate, source credibility, and spatiotemporal proximity. It innovatively introduces a gradient update based on sensitivity analysis and an adaptive time momentum smoothing mechanism, achieving online self-evolution of feature weights. Experimental results strongly validate the effectiveness of this strategy, demonstrating significant superiority over traditional static weighted and single-event chain models in both F1 score and early warning response speed. This successfully solves the challenge of balancing false alarms and missed alarms in complex public opinion environments.

[0139] Furthermore, this solution not only boasts superior algorithmic performance but also demonstrates exceptional business interpretability and practical value. Through visual analysis of the weight evolution trajectory, the system can clearly reveal the intrinsic driving factors of public opinion at different lifecycle stages (such as the outbreak, diffusion, and decline phases), assisting managers in shifting from "passive response" to "proactive analysis." Although dynamic calculation increases model complexity, relying on a distributed deployment architecture and optimized parameter settings, this method fully meets the stringent real-time requirements of tourism supervision scenarios.

[0140] Future work will focus on deepening understanding in three dimensions: First, expanding the boundaries of perception and verifying the model's adaptability when integrating more multidimensional modal data such as tourist physiological indicators and scenic area environmental monitoring; second, deepening graph analysis and optimizing the construction of dynamic semantic event chains by combining graph neural networks and deep causal inference technology to uncover deeper causal logic between events; and third, promoting application closed loops, exploring the linkage mechanism between early warning systems, intelligent customer service, and automated handling processes, and building a safe and efficient intelligent tourism governance ecosystem while strictly adhering to privacy regulations.

[0141] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0142] Figure 5 A structural block diagram of an example of a tourism public opinion event identification system based on dynamic weight adjustment according to an embodiment of this application is shown.

[0143] like Figure 5 As shown, the tourism public opinion event identification system 500 based on dynamic weight adjustment includes a data acquisition unit 510, a data preprocessing and structured representation unit 520, a key evaluation feature extraction unit 530, an event scoring and early warning unit 540, a feedback collection and accuracy evaluation unit 550, and a contribution calculation and weight update unit 560.

[0144] The data acquisition unit 510 is used to acquire multi-source tourism public opinion data associated with the target tourist destination. The multi-source tourism public opinion data includes at least public opinion content data and publication time information, data source identification information, publisher information and geographical location information associated with the public opinion content data. The public opinion content data includes text content and / or media content.

[0145] The data preprocessing and structured representation unit 520 is used to perform cleaning and structured processing on the multi-source tourism public opinion data, and map the public opinion content data into a structured feature representation with a unified dimension to generate a set of public opinion records. Each public opinion record in the set of public opinion records contains the structured feature representation and the corresponding publication time information, data source identification information, publisher information and geographical location information.

[0146] The key evaluation feature extraction unit 530 is used to extract multi-dimensional key evaluation features for event identification from each public opinion record in the public opinion record set based on a preset feature definition, so as to generate a corresponding key evaluation feature vector set.

[0147] The event scoring and early warning unit 540 is used to obtain the feature weight vector corresponding to the current monitoring period, and perform weighted combination processing on the key evaluation feature vector set based on the feature weight vector to obtain the event scoring result of each public opinion record; compare the event scoring result with the event judgment threshold to identify potential tourism public opinion events, and generate early warning information and record the event identifier associated with the early warning information if the potential tourism public opinion event is determined to exist.

[0148] The feedback acquisition and accuracy evaluation unit 550 is used to collect actual event evolution feedback data corresponding to the event identifier after the warning, and calculate the recognition accuracy based on the deviation between the actual event evolution feedback data and the warning information.

[0149] The contribution calculation and weight update unit 560 is used to calculate the contribution of each of the key evaluation features based on the recognition accuracy, and dynamically update the feature weight vector according to the contribution to obtain the updated feature weight vector for the next monitoring period; wherein, the key evaluation feature with a larger contribution is assigned a higher weight proportion in the updated feature weight vector.

[0150] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. The execution instructions can be read and executed by electronic devices (including but not limited to computers, servers, or network devices) to perform the steps of any of the above-described methods for identifying tourism public opinion events based on dynamic weight adjustment.

[0151] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described methods for identifying tourism public opinion events based on dynamic weight adjustment.

[0152] In some embodiments, this application also provides an electronic device, comprising: 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, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of a tourism public opinion event identification method based on dynamic weight adjustment.

[0153] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0154] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0155] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0156] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been 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.

Claims

1. A method for identifying tourism-related public opinion events based on dynamic weight adjustment, characterized in that, The method includes: Acquire multi-source tourism public opinion data associated with the target tourist destination. The multi-source tourism public opinion data includes at least public opinion content data and publication time information, data source identification information, publisher information and geographical location information associated with the public opinion content data. The public opinion content data includes text content and / or media content. The multi-source tourism public opinion data is cleaned and structured, and the public opinion content data is mapped to a structured feature representation with a unified dimension to generate a set of public opinion records. Each public opinion record in the set contains the structured feature representation as well as the corresponding publication time information, data source identification information, publisher information and geographical location information. Based on the preset feature definition, multi-dimensional key evaluation features for event identification are extracted from each public opinion record in the public opinion record set to generate a corresponding set of key evaluation feature vectors. Obtain the feature weight vector corresponding to the current monitoring period, and perform weighted combination processing on the key evaluation feature vector set based on the feature weight vector to obtain the event score result of each public opinion record; compare the event score result with the event judgment threshold to identify potential tourism public opinion events, and generate early warning information and record the event identifier associated with the early warning information if the potential tourism public opinion event is determined to exist; After the warning is issued, actual event evolution feedback data corresponding to the event identifier is collected, and the recognition accuracy is calculated based on the deviation between the actual event evolution feedback data and the warning information. The contribution of each key evaluation feature is calculated based on the recognition accuracy, and the feature weight vector is dynamically updated according to the contribution to obtain the updated feature weight vector for the next monitoring period; wherein, the key evaluation feature with a greater contribution is assigned a higher weight proportion in the updated feature weight vector.

2. The method according to claim 1, characterized in that, The step of cleaning and structuring the multi-source tourism public opinion data, and mapping the public opinion content data into a unified-dimensional structured feature representation, includes: The publisher information is subjected to statistically based abnormal information source filtering. The publishing frequency index and content duplication rate index within the preset sliding time window are calculated. Publishers whose index exceeds the preset abnormal threshold are marked as noise sources. The public opinion data published by the noise sources within the preset sliding time window are removed. The filtered text content is preprocessed by word segmentation and stop word removal, and non-standard dialect words in the text are identified by dialect correction model. Based on semantic similarity, the non-standard dialect words are mapped to standard semantic words. Then, a pre-trained language model is called to encode the processed text to generate text embedding vectors. Perform multi-channel feature extraction on the media content, including: calling an optical character recognition algorithm to extract scene text from image or video frames, and converting the scene text into auxiliary text vectors; The visual analysis model is invoked to extract visual content features of images or video frames to generate visual embedding vectors, and the speech emotion recognition model is invoked to extract acoustic emotion features from the video audio stream. A multimodal feature fusion space is constructed, and the text embedding vector, auxiliary text vector, visual embedding vector, and acoustic emotion features are projected onto a shared semantic dimension space using the corresponding learnable projection matrices. Normalization is performed on each modality vector after projection, and feature concatenation is performed to generate a structured feature representation of the corresponding public opinion record. .

3. The method according to claim 1, characterized in that, The process involves extracting multidimensional key evaluation features for event identification from each public opinion record in the set of public opinion records based on a preset feature definition, including: Regarding the first in the aforementioned set of public opinion records Each public opinion record is used to obtain its corresponding structured feature representation. And associated attribute information; Representing the structured features Input a pre-set sentiment analysis model for semantic decoding to calculate the sentiment intensity of the public opinion records. And calculate the influence based on the associated attribute information. ,growth rate Source credibility and spatiotemporal proximity In order to construct the corresponding multidimensional key evaluation feature vector; Among them, the influence The calculation is performed by constructing a weighted logarithmic interaction model, which uses a logarithmic function to smooth the publisher attribute data to suppress long-tailed distribution differences. ， In the formula, The number of followers listed in the publisher's information. The number of reposts of public opinion content within a preset observation window. The number of comments on public opinion content within the preset observation window; and These are the weighting coefficients for the breadth of dissemination and the weighting coefficients for the depth of interaction, respectively. Among them, the spatiotemporal proximity The calculation is performed by constructing a spatiotemporal Gaussian coupled decay model to quantify the degree to which public opinion records converge toward the core event in the spatiotemporal dimension. ， In the formula, The aforementioned geographical location information, As the core coordinates of the target tourist destination, This is a function for calculating geodesic distance; The publication time information, This refers to the sampling time of the current monitoring period or the end time of the current time window; This characterizes the decay of the timeliness of public opinion information. and These are the spatial dimension decay factor and the time dimension decay factor, respectively.

4. The method according to claim 3, characterized in that, The step of calculating the contribution of each of the key evaluation features based on the recognition accuracy, and dynamically updating the feature weight vector according to the contribution, includes: Based on the actual event evolution feedback data, the real event labels are determined, and a differentiable cross-entropy loss function is constructed in combination with the event scoring results to quantify the difference between the predicted distribution and the real distribution within the current monitoring period. Sensitivity analysis is performed on the feature weight vector by calculating the cross-entropy loss function with respect to the first... The partial derivatives of the weights of each key evaluation feature are used to determine the gradient contribution of that feature within the current monitoring period. The gradient contribution value characterizes the marginal utility of the feature weight change in reducing the prediction error; A multidimensional attention softmax mechanism based on exponential weighting is used to perform nonlinear mapping and update the feature weight vector to achieve adaptive weight allocation based on feature importance. ， Among them, the normalized partition function Defined as: ， In the formula, and These represent the current monitoring period. The Middle The first feature and the first The weight values ​​of each feature, Indicates the next monitoring cycle The Middle Update weights for each feature. The learning rate parameter is used to control the step size of the weight adjustment. A set of type identifiers for multidimensional key evaluation features. These represent emotional intensity, influence, growth rate, source credibility, and spatiotemporal proximity, respectively. Represents a set Any feature type index in the, Indicates the first The gradient contribution value corresponding to each feature type.

5. The method according to claim 4, characterized in that, After updating the feature weight vector using a multidimensional attention softmax mechanism based on exponential weighting through nonlinear mapping, the method further includes performing adaptive time momentum smoothing on the updated feature weight vector to suppress parameter oscillations, including: Establish a length of The sliding observation window is used to calculate the first... The temporal variance of gradient contribution values ​​within each monitoring period is used to construct a gradient instability index reflecting the system's evolution state. : ， In the formula, Features The mean within the sliding window, The total dimension of the features. Indicates the first The first monitoring cycle The gradient contribution value of each feature; the gradient instability index Used to quantify the degree of fluctuation in the contribution of features on a periodic sequence; Construct a nonlinear momentum adjustment function based on the gradient instability index. Dynamic calculation of the first Momentum smoothing coefficient for each monitoring period : ， In the formula, and These are the lower and upper bounds of the momentum smoothing coefficient, respectively. The preset stability threshold, The gain coefficient is used to adjust the response sensitivity; Based on the momentum smoothing coefficient Perform weighted smooth updates: ， In the formula, For the next monitoring cycle The Middle The final effective smoothing weights of each feature, The next monitoring period is obtained by using an exponentially weighted multidimensional attention softmax mapping. The Middle Update weights for each feature. For the current monitoring period The Middle Historical smoothing weights for each feature.

6. The method according to claim 5, characterized in that, The method further includes constructing a dynamic semantic event chain graph using smoothed feature weight vectors processed by adaptive time momentum smoothing to track the evolution path of public opinion, including: The potential tourism-related public opinion events are clustered into a set of event nodes, and for any two event nodes... and , using the Smoothing weights for each monitoring period Calculate event nodes and Cross-feature weighted cosine similarity between : ， In the formula, and Representing event nodes respectively and In the Aggregated feature values ​​on class features A set of type identifiers for multidimensional key evaluation features. For the next monitoring cycle The Middle The final effective weights for each feature are smoothed; the similarity formula is applied through weights. The feature space is dynamically stretched so that features with high weight in the current period play a dominant role in similarity measurement. Based on the cross-feature weighted cosine similarity Construct a directed acyclic graph and assign similarity Event nodes that exceed a preset link threshold are sequentially connected to generate the event chain graph.

7. The method according to claim 6, characterized in that, After generating the event chain graph, the method further includes extracting key public opinion evolution paths and identifying key points of public opinion mutation based on graph topology analysis, including: Define event nodes in the event chain graph Cumulative path energy state A state transition equation is constructed using a dynamic programming algorithm to search for the globally optimal main public opinion evolution path. : ， In the formula, Represents all pointing nodes in the directed acyclic graph. The set of predecessor nodes, express Any node in, For event nodes The cumulative path energy state, For nodes The influence characteristic value, Path energy adjustment factor; Represents event nodes and Cross-feature weighted cosine similarity between them; By backtracking The sequence of nodes that reaches the global maximum value determines the main public opinion evolution path. , The first in the main public opinion evolution path 1 node The total number of nodes in the main public opinion evolution path; Along the aforementioned main public opinion evolution path Perform sentiment gradient detection and calculate the discrete sentiment gradient between adjacent nodes. : ， In the formula, and The first on the path The node and the first The emotional intensity feature value of each node, The time interval between events corresponding to the two nodes. The preset regularization parameter is used to avoid zero denominators and improve the numerical stability of gradient calculation; it will satisfy... nodes Identified as a key point of sudden change in public opinion, among which This is the preset mutation detection threshold.

8. The method according to claim 5, characterized in that, The method further includes adaptively optimizing the event determination threshold based on the F1 score maximization criterion, including: Utilize the next monitoring cycle Smoothing weights Recalculate the event scores of the historical validation set samples and iterate through multiple candidate thresholds within the preset interval; For each candidate threshold, the corresponding number of true positives is counted. False positives and false negatives And calculate the accuracy. and recall rate : ， Based on the accuracy and recall rate Calculate the harmonic mean index F1 score : ， Choose the option that makes the F1 score The candidate threshold that is maximized is set as the next monitoring period. The event judgment threshold is set to achieve a dynamic balance between false alarm rate and false negative rate.

9. The method according to claim 3, characterized in that, Regarding the growth rate The calculations include: Using the exponential moving average model for the first The number of mentions of the sentiment topic or destination entity associated with each sentiment record at the current sampling time. First-order difference rate Perform smoothing calculations: ， In the formula, For the public opinion topic or destination entity at the previous sampling time Smooth growth rate This is the attenuation smoothing coefficient; Regarding emotional intensity The calculations include: Construct an attention-based semantic decoding model for the structured feature representation. A joint analysis of emotional polarity and intensity: ， In the formula, This is a self-attention mechanism function used to capture long-distance semantic dependencies within feature vectors; and These are the weight matrix and bias term of the semantic decoding model, respectively; It is a hyperbolic tangent activation function, whose absolute output value represents the intensity of emotion from calm to agitation.

10. A tourism public opinion event identification system based on dynamic weight adjustment, characterized in that, The system includes: The data acquisition unit is used to acquire multi-source tourism public opinion data associated with the target tourist destination. The multi-source tourism public opinion data includes at least public opinion content data and publication time information, data source identification information, publisher information and geographical location information associated with the public opinion content data. The public opinion content data includes text content and / or media content. The data preprocessing and structured representation unit is used to perform cleaning and structured processing on the multi-source tourism public opinion data, and map the public opinion content data into a structured feature representation with a unified dimension to generate a set of public opinion records. Each public opinion record in the set of public opinion records contains the structured feature representation and the corresponding publication time information, data source identification information, publisher information and geographical location information. The key evaluation feature extraction unit is used to extract multi-dimensional key evaluation features for event identification from each public opinion record in the public opinion record set based on a preset feature definition, so as to generate a corresponding key evaluation feature vector set. The event scoring and early warning unit is used to obtain the feature weight vector corresponding to the current monitoring period, and perform weighted combination processing on the key evaluation feature vector set based on the feature weight vector to obtain the event scoring result of each public opinion record; compare the event scoring result with the event judgment threshold to identify potential tourism public opinion events, and generate early warning information and record the event identifier associated with the early warning information if the existence of the potential tourism public opinion event is determined; The feedback collection and accuracy evaluation unit is used to collect actual event evolution feedback data corresponding to the event identifier after the warning, and calculate the recognition accuracy based on the deviation between the actual event evolution feedback data and the warning information; The contribution calculation and weight update unit is used to calculate the contribution of each of the key evaluation features based on the recognition accuracy, and dynamically update the feature weight vector according to the contribution to obtain the updated feature weight vector for the next monitoring period; wherein, the key evaluation feature with a larger contribution is assigned a higher weight proportion in the updated feature weight vector.

Images