Lightweight large model deployment method suitable for resource-constrained environment

Abstract

The invention discloses a lightweight large model deployment method suitable for a resource-constrained environment, and relates to the technical field of large model lightweight, and the method comprises the following steps: comprehensively collecting and obtaining all potential public opinion topics through an API interface of a social media platform or a crawler technology, constructing a basic data set of public opinion analysis, and constructing a basic data set of public opinion analysis; and public opinion trends on the current social media are comprehensively grasped. According to the method, classification processing and deep analysis are carried out on public opinion topics of the social media platform, implicit or private languages in the public opinion topics are recognized in real time, when the occurrence frequency of the languages is increased, the analysis threshold value is adjusted in a self-adaptive mode, the sensitivity to potential social unstable factors is enhanced, and therefore misjudgment and missing of intervention opportunities are avoided, and user experience is improved. Therefore, the police service department can timely and accurately identify and deal with potential social threats in a resource-limited environment, and social fluctuation or violence events can be effectively prevented in an initial stage.

Description

Technical Field

[0001] The present invention relates to the technical field of large model lightweighting, and specifically relates to a lightweight large model deployment method applicable to resource-constrained environments. Background Art

[0002] The lightweighting of large models for deployment in resource-constrained environments aims to enable large models that are originally huge, have high computing requirements, and consume a large amount of resources to operate efficiently in devices or environments with limited computing power and storage through various technical means and optimization strategies. In the field of police practical operations, this optimization can ensure that in tasks such as video surveillance, real-time analysis, and intelligent patrol, lightweight artificial intelligence models can be deployed on low-power, resource-constrained terminal devices, thereby achieving efficient intelligence processing and rapid identification and response to abnormal events. The core technologies of lightweighting include model compression, knowledge distillation, quantization, pruning, etc. By reducing model parameters, lowering computational complexity, and compressing the model volume, the flexibility and response speed of deployment are improved while ensuring performance.

[0003] In the field of police practical operations, the role of social public opinion monitoring and analysis is crucial. Especially in the context of the lightweighting of large models for deployment in resource-constrained environments, its advantages can be fully exerted. By real-time monitoring and analyzing public opinion information on platforms such as social media, news reports, and forum discussions, the police department can timely discover potential social instability factors and public safety risks, such as the incitement of violent events, the expansion of rallies and protests, and the public opinion storm of sudden social events. Using lightweight large models, massive data can be quickly processed on resource-limited terminal devices, key information can be captured in real time, the development trend of public opinion can be predicted, and timely intelligence support can be provided to decision-makers, helping police officers make effective responses at the initial stage of the development of the situation, avoiding the deterioration of the situation, and thus ensuring social harmony and stability.

[0004] The prior art has the following deficiencies: During the process of social public opinion monitoring, with the lightweighting of large models, the reduction of model parameters and computational complexity may reduce the ability to identify subtle public opinion signals, thereby leading to misjudgments. Especially some implicit or indirect social media information, such as suggestive or veiled language expressions, may be ignored or misunderstood by the lightweighted model, resulting in the missed potential social instability risks. Such misjudgments may prevent the police department from timely identifying real threats, missing the best intervention opportunity, and even leading to the aggravation of social unrest or violent events.

[0005] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and thus it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0006] The object of the present invention is to provide a lightweight large model deployment method applicable to resource-constrained environments. By classifying and deeply analyzing the public opinion topics on social media platforms, it can dynamically monitor topic changes and real-time identify suggestive or implicit language. When the frequency of such language increases, it adaptively adjusts the analysis threshold to enhance the sensitivity to potential social instability factors, thereby avoiding misjudgment and missing the opportunity for intervention. Through this lightweight large model deployment method applicable to resource-constrained environments, not only the accuracy and efficiency of public opinion monitoring are improved, but also it ensures that in resource-constrained environments, the police department can timely and accurately identify and respond to potential social threats, thereby effectively preventing the occurrence of social unrest or violent incidents in the initial stage to solve the problems in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions: A lightweight large model deployment method applicable to resource-constrained environments, including the following steps: Through the API interface or crawler technology of social media platforms, comprehensively collect and obtain all potential public opinion topics, construct a basic data set for public opinion analysis, and comprehensively grasp the current public opinion dynamics on social media; Classify the collected public opinion data according to topics, and establish an analysis set for each public opinion topic separately, so that each public opinion topic is processed independently; During the analysis process, for each public opinion topic, screen out the text data containing suggestive or implicit language, deeply analyze the screened data, and input the data after in-depth analysis into the pre-trained deep learning model to evaluate the frequency of suggestive and implicit language in each public opinion topic; Dynamically monitor the changes of public opinion topics through the deep learning model, and real-time track whether a certain topic frequently appears suggestive or implicit language. When it is found that the frequency of suggestive and implicit language in a certain public opinion topic increases, it will automatically trigger a threshold relaxation mechanism to adaptively adjust the analysis threshold and enhance the sensitivity to suggestive and implicit language.

[0008] Preferably, the public opinion topics are divided based on keywords, popular tags, event names or regions, which are used to accurately identify and locate the focus of attention on current social media.

[0009] Preferably, the deep learning model is a natural language processing model based on the Transformer architecture.

[0010] Preferably, for each public opinion topic, screen out the text data containing suggestive or implicit language through the following steps: Utilize the sentiment analysis model in natural language processing technology to extract suggestive language expressions by identifying the implicit emotions or feelings in the text; Capture words and phrases with potential implications or veiled expressions using keyword matching and semantic analysis; Apply context understanding techniques to determine whether public opinion statements contain implicit or suggestive content in combination with context information; Combine sentiment lexicons with semantic role labeling, compare the sentiment tendencies of words with their usage in the context, and further identify the language features of implicit or suggestive content.

[0011] Preferably, conduct in-depth analysis on the screened data, input the data after in-depth analysis into a pre-trained deep learning model, and evaluate the occurrence frequency of suggestive and veiled language in each public opinion topic. The specific steps are as follows: Extract features reflecting the high-frequency occurrence of suggestive and veiled language from the screened data. Among them, the extracted features include the frequency of vague or ambiguous language expressions in the text data and the frequency of conveying information in a non-literal way in the text data. After in-depth analysis of the screened data, generate a reference value for the ratio of vague expressions and a reference value for non-literal communication respectively. Through the reference value for the ratio of vague expressions and the reference value for non-literal communication, initially quantify the occurrence frequency of suggestive and veiled language in the text data and its potential emotional intensity to help identify potential social instability risks; Input the reference value for the ratio of vague expressions and the reference value for non-literal communication into a pre-trained deep learning model, generate a frequency coefficient through the model, and evaluate the occurrence frequency of suggestive and veiled language in each public opinion topic based on the frequency coefficient.

[0012] Preferably, the specific steps for generating a reference value for the ratio of vague expressions after in-depth analysis of the frequency of vague or ambiguous language expressions in the text data are as follows: In the text data, first identify the ambiguity of all language expressions through natural language processing techniques. For each piece of text data, use an ambiguity vocabulary library and a model based on deep semantic parsing to identify the ambiguous language expressions therein, and assign a weighted score to the ambiguous words. The specific assignment formula is: , where is the weighted score for the ambiguous expression in the text ; is the ambiguity score of the word w in the ambiguity vocabulary library, reflecting the ambiguity degree of the word; is the set of semantic roles of the word w in the context; is the context weight of the semantic role c , calculated based on the deep semantic parsing model, used to reflect the specific meaning and potential emotion of the word in this context; After the text is scored for ambiguity, the next step is to calculate the reference value of the ambiguous expression ratio. The overall ambiguous expression ratio is obtained by summarizing the ambiguity scores of all texts, further quantifying the frequency of suggestive and implicit language and its potential emotional intensity in this topic. The specific quantification formula is as follows: , where, is the reference value of the ambiguous expression ratio for the topic T ; is the set ambiguity threshold for screening out important ambiguous language; represents the number of texts that meet the conditions, that is, the number of texts with significant ambiguous expressions.

[0013] Preferably, after in-depth analysis of the frequency of information transmitted in a non-literal way in the text data, the specific steps to generate the non-literal communication reference value are as follows: When analyzing the frequency of information transmitted in a non-literal way in the text, first, an identification model for non-literal communication needs to be established. This identification model is based on natural language processing technology and combines multiple semantic analysis methods to extract the characteristics of suggestive, implicit, and indirect language, and assigns a non-literal communication weight to the non-literal expression. The calculation expression of the non-literal communication weight is as follows: , where, represents the non-literal communication weight in text t, is the non-literal expression intensity function of text t in the context change, is the weight parameter of each non-literal language feature, is the total length or total time period of the text; After obtaining the non-literal weight of each text, the obtained non-literal weights are normalized and combined to generate the non-literal communication reference value. The generated expression is as follows: , where, is the non-literal communication reference value, is the adjustment factor of each text segment.

[0014] Preferably, when it is found that the frequency of suggestive and implicit language in a certain public opinion topic increases, the threshold relaxation mechanism is automatically triggered to adaptively adjust the analysis threshold. The specific steps are as follows: When the frequency coefficient exceeds the frequency coefficient reference threshold , the threshold relaxation mechanism is automatically started to adjust the analysis threshold. The new analysis threshold is adjusted by the following formula, and the adjustment expression is: , Among them, is the current analysis threshold, is the adjusted analysis threshold; After the threshold is relaxed, the new analysis threshold is applied to the deep learning model and follows the following formula: , If , trigger the second round of threshold adjustment; otherwise, restore the initial threshold .

[0015] In the above technical solution, the technical effects and advantages provided by the present invention are as follows: By classifying and deeply analyzing the public opinion topics on the social media platform, the present invention can dynamically monitor the topic changes and real-time identify the suggestive or implicit language therein. When the frequency of such language increases, it adaptively adjusts the analysis threshold to enhance the sensitivity to potential social instability factors, thereby avoiding misjudgment and missing the intervention opportunity. Through this lightweight large model deployment method applicable to resource-constrained environments, it not only improves the accuracy and efficiency of public opinion monitoring, but also ensures that in resource-constrained environments, the police department can timely and accurately identify and respond to potential social threats, thereby effectively preventing the occurrence of social unrest or violent incidents in the initial stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0017] Figure 1 is the method flow chart of the lightweight large model deployment method applicable to resource-constrained environments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Now the example embodiments will be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the example embodiments to those skilled in the art.

[0019] The present invention provides a lightweight large model deployment method applicable to resource-constrained environments as shown in Figure 1 and includes the following steps: Through the API interface or web crawler technology of social media platforms, comprehensively collect and obtain all potential public opinion topics, construct a basic dataset for public opinion analysis, and comprehensively grasp the current public opinion dynamics on social media; These public opinion topics can be based on specific keywords, popular hashtags, event names, or regional divisions. Classifying public opinion topics based on specific keywords, popular hashtags, event names, or regional divisions can help accurately identify and locate the focus of attention on current social media, ensuring the comprehensiveness and pertinence of public opinion analysis. Through the classification of keywords and hashtags, discussions related to specific topics can be effectively screened out, avoiding information overload, and at the same time concentrating resources for in-depth analysis of relevant topics. Classification based on event names or regions helps to track and monitor emergencies or local public opinions in real time, enables timely identification of potential threatening social dynamics, and allows for a quick response and corresponding measures. Such a classification method improves the analysis efficiency, avoids interference from a large amount of irrelevant data, and enhances the accuracy of public opinion early warning. The data obtained usually includes text content, comments, reposts, etc. from platforms such as Weibo, WeChat, Twitter, etc.

[0020] Classify the collected public opinion data according to topics, and create an analysis set for each public opinion topic separately, so that each public opinion topic can be processed independently; The data in the analysis set will include all text content, sentiment tags, timestamps, user behaviors, etc. related to the topic. The role of this step is to clearly separate public opinion topics at the data level, enabling each topic to be processed independently and avoiding information interference. By creating an analysis set for each topic separately, the public opinion dynamics of a certain topic can be accurately tracked and analyzed, effectively reducing data complexity and improving analysis efficiency.

[0021] During the analysis process, for each public opinion topic, screen out the text data containing suggestive or implicit language, conduct in-depth analysis on the screened data, and input the data after in-depth analysis into a pre-trained deep learning model to evaluate the occurrence frequency of suggestive and implicit language in each public opinion topic; For each public opinion topic, screen out the text data containing suggestive or implicit language through the following steps: First, use the sentiment analysis model in natural language processing (NLP) technology to extract suggestive language expressions by identifying the implicit emotions or feelings in the text; Second, use keyword matching and semantic analysis to capture words and phrases with potential suggestions or implicit expressions, such as by identifying special expressions such as puns, irony, metaphors, etc.; Third, apply the technology of context understanding, and combine context information to judge whether public opinion words contain implicit or suggestive content; Finally, by combining the sentiment dictionary with semantic role labeling, compare the sentiment tendency of words with their usage in the context to further identify the language features containing implicit or suggestive content.

[0022] Through these methods, potential suggestive and implicit language can be efficiently and accurately screened out, providing deeper insights for public opinion analysis.

[0023] Sentiment analysis is one of the most common techniques in natural language processing. By analyzing sentiment words, sentence structures, tones, etc. in the text, it helps to identify potential sentiment tendencies. In this process, special attention needs to be paid to texts that do not directly express negative emotions or attitudes, as these texts often express dissatisfaction, criticism, or potential social unrest in an implicit way. For example, some seemingly neutral or positive comments may show potential reverse emotions through subtle sentiment words and tones. Sentiment analysis models can not only extract the intuitive sentiment in the text but also use machine learning to identify those indirect or implicit sentiment signals, which is crucial for capturing suggestive language. Through this process, texts that may contain suggestive or implicit sentiment can be quickly screened out from a large amount of social media data, providing a basis for subsequent analysis.

[0024] Keyword matching mainly detects the occurrence frequency of specific words through a pre-set keyword library to help capture keywords or phrases with potential suggestions or implicit meanings. Especially when social media topics are related to sensitive, political, or highly emotional events, users often use some seemingly harmless or common words to express themselves. These words themselves do not have intuitive aggression or negative sentiment, but through the role of context and context, they may convey potential negative emotions. To more accurately identify such content, semantic analysis techniques are needed to compare the meaning of keywords with the context to further infer their potential meanings. For example, language forms such as puns, irony, and metaphors are particularly common in social media. These forms of expression may imply the author's deep dissatisfaction with an event or phenomenon. Semantic analysis can effectively capture these special language phenomena by understanding the polysemy of words and their specific meanings in a specific context, thus improving the ability to identify implicit expressions.

[0025] In social media platforms, a single word often fails to reveal the true meaning of a text. Many veiled or suggestive languages rely on the overall context for understanding. For example, a seemingly ordinary sentence like "Be careful" might be misjudged as neutral or harmless without context. However, when combined with the context (such as subsequent mentions of violence, conflicts, etc.), it may reveal its hidden potential warning or sentiment. To effectively capture such meaning, context analysis techniques are needed to comprehensively analyze the sentence with the surrounding text and understand the implied meaning. This approach can help the model delve deeper into the potential risks in the text. Especially when dealing with complex social dynamics or sensitive topics, context understanding techniques are of particular importance.

[0026] Finally, the combination of sentiment lexicons and semantic role labeling provides important support for further improving the recognition of suggestive language. A sentiment lexicon is a collection of words with a large number of sentiment annotations. It can not only provide the sentiment tendencies of common sentiment words but also support the sentiment meanings in some complex contexts. In public opinion analysis, sentiment lexicons help identify those implied or veiled sentiment information. Especially on social media, many users avoid direct negative emotions when using sentiment words and instead express dissatisfaction in a more indirect way. Through sentiment lexicons, these complex sentiment changes can be accurately captured, and thus the potential implications in the text can be identified. Parallel to this is semantic role labeling, which helps analyze the key roles and actions in a sentence by identifying the grammatical structure and the relationships between words in the sentence. Semantic role labeling can help better understand the potential meaning in the text, such as who is the initiator of the action, who is the victim or beneficiary, which is particularly important for revealing veiled social emotions and sentiments.

[0027] Conduct in-depth analysis on the selected data, input the data after in-depth analysis into the pre-trained deep learning model, and evaluate the occurrence frequency of suggestive and veiled language in each public opinion topic. The specific steps are as follows: Extract the features reflecting the high-frequency occurrence of suggestive and veiled language from the selected data. Among them, the extracted features include the frequency of vague or ambiguous language expressions in the text data and the frequency of conveying information in a non-literal way in the text data. After in-depth analysis of the selected data, generate the reference values of the fuzzy expression ratio and the non-literal communication reference value respectively. Through the fuzzy expression ratio reference value and the non-literal communication reference value, initially quantify the occurrence frequency of suggestive and veiled language in the text data and its potential sentiment intensity to help identify potential social instability risks; Input the fuzzy expression ratio reference value and the non-literal communication reference value into the pre-trained deep learning model, generate a frequency coefficient through the model, and evaluate the occurrence frequency of suggestive and veiled language in each public opinion topic based on the frequency coefficient.

[0028] A pre-trained deep learning model refers to a deep neural network model that has been trained on a large and diverse set of public opinion datasets. After sufficient training, it can effectively analyze and predict new and unseen public opinion data. Such models usually rely on large datasets and learn deep features such as language patterns, sentiment expressions, and context analysis through methods like transfer learning or self-supervised learning. Their advantage lies in their ability to automatically extract high-dimensional features from historical data and transform these features into knowledge that can handle practical applications. A pre-trained deep learning model can identify implicit signals in text, including sentiment tendencies, language styles, and context relationships, thus providing more accurate public opinion analysis. In this way, the model can effectively capture implicit, indirect, and veiled language and convert this information into quantifiable metrics, such as the reference value of the vague expression ratio and the reference value of non-literal communication.

[0029] Pre-trained deep learning models play a crucial role in public opinion monitoring, especially when dealing with complex natural language. Specifically, such models can identify language patterns, sentiment tendencies, and potential information in text through multi-layer neural network architectures, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs). In public opinion topic analysis, pre-trained models can identify implicit or difficult-to-intuitively-understand language features, which is particularly important for accurately evaluating the implicit or veiled language in public opinion topics. For example, the model can understand the potential sentiment expression of specific words or sentences in different contexts through context information, thereby distinguishing between surface and deep meanings. By calculating metrics such as the reference value of the vague expression ratio and the reference value of non-literal communication, the model can output a frequency coefficient that indicates the frequency of the occurrence of veiled or implicit language in a certain public opinion topic and provides a reliable basis for subsequent decision-making and actions. Therefore, pre-trained deep learning models not only enhance the ability to process complex language but also improve the accuracy and response speed of public opinion monitoring systems, enabling police departments to take effective intervention measures earlier when facing potential threats.

[0030] A high frequency of vague or ambiguous language expressions in language texts usually indicates a high frequency of suggestive and implicit language in public opinion topics. In social media and online discussions, users often express sensitive or potential emotions through vague language, especially when facing complex or controversial topics. Vague language is usually not as explicit as directly expressing emotions. It often relies on means such as suggestion, metaphor, or irony to convey information, so it is difficult to capture through traditional sentiment analysis models. For example, expressions like "The situation is a bit complicated" or "It's hard to say about this matter" seem neutral or vague on the surface, but often conceal negative emotions or social anxiety. If similar vague or ambiguous language appears frequently in public opinion topics, it means that there may be a large number of implicit emotions or potential instability factors in the topic. Therefore, by monitoring the frequency of these vague languages, the implicit risks in public opinion can be effectively identified, especially in situations where it is difficult to express through straightforward language.

[0031] The specific steps to generate a reference value for the fuzzy expression ratio after in-depth analysis of the frequency of vague or ambiguous language expressions in text data are as follows: In text data, first, natural language processing (NLP) techniques are needed to identify the ambiguity of all language expressions. Vague expressions usually include non-straightforward and ambiguous wording, such as "pun", "metaphor", "irony", "indirect expression", etc. For each piece of text data, a fuzzy lexicon and a model based on deep semantic parsing are used to identify the language expressions that may contain ambiguity, and a weighted score is assigned to these fuzzy words. The formula is: , where, is the weighted score of the fuzzy expression in text ; is the ambiguity score of the word w in the fuzzy lexicon, reflecting the ambiguity degree of the word; is the set of semantic roles of the word w in the context; is the context weight of the semantic role c , calculated based on the deep semantic parsing model, used to reflect the specific meaning and potential emotion of the word in this context; The ambiguity score can be calculated through a variety of existing algorithms. The most common methods include calculations based on word vector models, sentiment analysis, and context analysis. First, word embedding algorithms (such as Word2Vec, GloVe, or FastText) can be used to obtain the ambiguity score by calculating the semantic distance between words. Through these models, the semantics of words are mapped to a high-dimensional vector space, where words with stronger ambiguity are usually associated with multiple meanings or emotions, manifested as a higher degree of vector dispersion. Second, sentiment analysis algorithms can be used to evaluate the emotional intensity and tendency conveyed by words in different contexts. For example, certain words may have different emotional colors in different contexts, and sentiment analysis can help identify these variations and provide a basis for the ambiguity score. Finally, context-dependent analysis (such as based on pre-trained models like BERT, GPT, etc.) can further refine the ambiguity score according to the usage of words in specific contexts. In complex sentences, the meaning of certain words may become ambiguous, and at this time, context analysis can infer the potential ambiguity of the word through the context to ensure the accuracy of the score. Combining these methods can comprehensively reflect the degree of vagueness of words, thus providing an accurate calculation basis for the ambiguity score.

[0032] The role of this step is to identify the ambiguous language in the text and accurately calculate the emotional potential and degree of ambiguity of this ambiguous language through a weighting mechanism. By assigning a weighted score to each ambiguous word, it ensures that the model can capture more delicate expressions and potential emotional signals.

[0033] After scoring the ambiguity of the text, the next step is to calculate the reference value of the ambiguous expression ratio. At this time, it is necessary to summarize the ambiguity scores of all texts to obtain the overall ambiguous expression ratio, further quantifying the frequency of occurrence of suggestive and implicit language in this topic and its potential emotional intensity. The formula is: , where, is the reference value of the ambiguous expression ratio for the topic T ; is the set ambiguity expression threshold used to screen out important ambiguous language (i.e., texts with a score greater than ); represents the number of texts that meet the conditions, that is, the number of texts with significant ambiguous expressions.

[0034] From the fuzzy expression ratio reference value, it can be seen that the larger the performance value of the fuzzy expression ratio reference value generated after in-depth analysis of the frequency of fuzzy or ambiguous language expressions in text data, the higher the frequency of suggestive and implicit language in the public opinion topic. This is because fuzzy expressions usually refer to those languages that are not direct, vague, and carry implicit emotions or attitudes. For example, emotions expressed through puns, irony, metaphors, etc. are often hidden behind the surface language and are not easily directly recognized. When the frequency of fuzzy language in the text is high, it means that in the discussion of this topic, more emotions and information are conveyed in an indirect way, often an indirect expression of complex, sensitive or controversial topics. The fuzzy expression ratio reference value can effectively quantify the accumulation of implicit emotions in the topic by measuring the frequency and weighted scores of these fuzzy expressions. When the performance value of the fuzzy expression ratio reference value is high, it usually indicates that the words in the topic are relatively vague, and the suggestive and implicit emotional components account for a larger proportion, which means that there are more potential social risks or emotional tensions in this topic.

[0035] A high frequency of information transmission in a non-straightforward manner in text data usually indicates a high frequency of suggestive and implicit language in this public opinion topic. In social media and public discussions, users often use implicit and indirect ways to express emotions or positions, especially when facing sensitive topics. Through non-straightforward language forms such as metaphors, puns, irony, symbolism, etc., users can avoid direct conflicts or evade censorship while conveying potential emotions or viewpoints. Therefore, if the frequency of such non-straightforward expressions is high in a certain public opinion topic, it usually means that the topic is filled with implicit or suggestive language, and these implicit signals may hide important social emotions or unstable factors, which deserve further attention and analysis.

[0036] The specific steps for generating the non-straightforward communication reference value after in-depth analysis of the frequency of information transmission in a non-straightforward manner in text data are as follows: When analyzing the frequency of information transmission in a non-straightforward manner in the text, first, a recognition model for non-straightforward communication needs to be established. This model is based on natural language processing (NLP) technology and combines multiple semantic analysis methods to extract the characteristics of suggestive, implicit, and indirect language. This process involves in-depth parsing of all words, phrases, and sentences in the text to identify whether there are non-straightforward languages such as puns, irony, sarcasm, symbolic expressions, or implicit tones. These non-straightforward expressions will be assigned a non-straightforward communication weight, which reflects the intensity of the suggestive or implicit content contained in this language segment. This non-straightforward communication weight is determined by multiple factors, including context, emotional relevance, language characteristics, etc. For this purpose, an integral model is constructed, and by weighted summing the non-straightforward characteristics of each text segment, the non-straightforward communication weight of each text is obtained. The formula is as follows: , wherein, represents the non-literal communication weight in text t, is the non-literal expression intensity function of text t in the context change, is the weight parameter of each non-literal language feature, is the total length or total time period of the text; The non-literal expression intensity function can be modeled by various existing functional forms, mainly depending on the semantic features, emotional features, and language structure in the text. A common choice is a sentiment analysis-based model, which quantifies the intensity of non-literal expressions by calculating the sentiment intensity and sentiment turning points in the text. For example, a Gaussian function can be used to represent the smooth change of sentiment fluctuations, and the shape of the Gaussian function can capture the concentrated areas of sentiment in the text. Especially when certain implicit sentiment expressions are concentrated in specific sentences or paragraphs, it can effectively weight the influence of these sentiments. In addition, the Sigmoid function can also be used to represent the gradual enhancement process of certain suggestive expressions in the text, especially suitable for non-linear and progressive sentiment expressions, such as gradually intensifying sarcasm or irony. The exponential function can reflect the sharp change of sentiment intensity or the expression of sudden emotions, and is suitable for phrases or paragraphs with strong implicit emotions. These functions can precisely capture the intensity of non-literal language expressions in the text through different mathematical forms, combined with specific weight parameters , further improving the accuracy of the analysis.

[0037] The role of this step is to precisely quantify the implicit or indirect expressions in each segment of the text by weighted integration of the non-literal language features in the text.

[0038] After obtaining the non-literal weights of each text, these weights are normalized and combined to generate a non-literal communication reference value. The purpose of this step is to transform the non-literal communication features in the text into a scaling index, reflecting the frequency of non-literal language and its potential sentiment intensity in this topic. Specifically, the non-literal communication reference value is the weighted sum of the non-literal weights of all text segments, considering the importance of different text segments (for example, long texts and segments with strong emotions have higher weights). The formula is as follows: , wherein, is the non-literal communication reference value, is the adjustment factor for each text segment, used to weight long texts, segments with high emotional intensity, etc. This adjustment factor reflects the importance or influence of the text segment. For example, for a text segment with high emotional intensity, its emotional tendency intensity can be calculated through a sentiment analysis model, and this intensity is combined with Multiply them to give these segments higher weights. Finally, the non-literal communication reference value is the cumulative intensity of non-literal language expressions in all texts and can reveal the potential emotional trends of suggestive and implicit language in a topic.

[0039] From the non-literal communication reference value, the larger the performance value of the non-literal communication reference value generated after in-depth analysis of the frequency of information transmitted in a non-literal manner in the text data, usually means that the frequency of suggestive and implicit language in the public opinion topic is higher. The non-literal communication reference value is generated by in-depth analysis of the frequency and emotional intensity of information transmitted in a non-literal manner in the text. It can quantify the language that expresses emotions and positions indirectly through metaphors, puns, irony, sarcasm, etc. When non-literal language appears frequently, the non-literal communication reference value of the text will increase accordingly, reflecting the concealment and complexity of the potential emotions or social signals in the topic. Therefore, the increase in the non-literal communication reference value not only indicates the frequent occurrence of implicit language in the topic, but may also reveal deeper social emotions or potential instability factors behind it, which deserve special attention.

[0040] The deep learning model is not specifically limited here. Any deep learning model that can realize the comprehensive analysis of the fuzzy expression ratio reference value and the non-literal communication reference value to generate the frequency coefficient is acceptable. To implement the technical solution of the present invention, the present invention provides a specific implementation; the expression for generating the frequency coefficient is: , in the formula, , are respectively the preset proportionality coefficients of the fuzzy expression ratio reference value and the non-literal communication reference value , and , are both greater than 0. According to the given content, the preset proportionality coefficient refers to the coefficient preset in the deep learning model for adjusting the weights between different variables. In the formula, and are respectively two parameter coefficients of the model, used to adjust the relationship between the two main variables of the fuzzy expression ratio reference value and the non-literal communication reference value in the model. According to this formula, the preset proportionality coefficients and are used to affect the weighted contributions of each part in the model, and are usually adjusted and optimized through experiments or historical data to ensure that the model can better adapt to the actual data and generate accurate predictions.

[0041] As can be seen from the frequency coefficient, the larger the performance value of the reference value of the fuzzy expression ratio generated after in-depth analysis of the frequency of fuzzy or ambiguous language expressions in the text data, the larger the performance value of the reference value of non-literal communication generated after in-depth analysis of the frequency of information transmitted in a non-literal way in the text data. That is, the larger the performance value of the frequency coefficient generated when the deep learning model completed through pre-training predicts the frequency of suggestive and implicit language in each public opinion topic, the higher the frequency of suggestive and implicit language in the public opinion topic. Conversely, it indicates that the frequency of suggestive and implicit language in the public opinion topic is lower.

[0042] Compare and analyze the frequency coefficient generated when the deep learning model completed through pre-training predicts the frequency of suggestive and implicit language in each public opinion topic with the preset reference threshold of the frequency coefficient to evaluate the frequency of suggestive and implicit language in each public opinion topic. The specific steps are as follows: If the frequency coefficient is greater than the reference threshold of the frequency coefficient, the frequency of suggestive and implicit language in this public opinion topic is classified as frequently occurring; if the frequency coefficient is less than or equal to the reference threshold of the frequency coefficient, the frequency of suggestive and implicit language in this public opinion topic is classified as normally occurring.

[0043] Dynamically monitor the changes in public opinion topics through the deep learning model, and real-time track whether a certain topic frequently appears suggestive or implicit language. When it is found that the frequency of suggestive and implicit language in a certain public opinion topic increases, the threshold relaxation mechanism will be automatically triggered to adaptively adjust the analysis threshold and enhance the sensitivity to suggestive and implicit language; When it is found that the frequency of suggestive and implicit language in a certain public opinion topic increases (that is, the frequency of suggestive and implicit language in this public opinion topic is classified as frequently occurring), the threshold relaxation mechanism will be automatically triggered to adaptively adjust the analysis threshold. The specific steps are as follows: The frequency coefficient of each topic is compared with the preset reference threshold of the frequency coefficient The reference threshold of the frequency coefficient is set initially and is used to indicate the normal range of the frequency of suggestive or implicit language in the public opinion topic. If the frequency coefficient is greater than the reference threshold of the frequency coefficient , it means that the implicit language in this topic frequently appears, and there may be potential social instability risks. At this time, the system will identify this change and trigger the threshold relaxation mechanism to further adjust the analysis threshold to enhance the sensitivity to the suggestive language in this topic; The core purpose of the threshold relaxation mechanism is to improve the sensitivity of the model to frequently occurring implicit signals by adjusting the analysis threshold, so as to avoid missing potential risks.

[0044] When the frequency coefficient exceeds the frequency coefficient reference threshold , the system will automatically activate the threshold relaxation mechanism to adjust the analysis threshold. The new analysis threshold is adjusted by the following formula, and the adjustment expression is: , where is the current analysis threshold, is the adjusted analysis threshold; Through this formula, as the frequency coefficient exceeds the frequency coefficient reference threshold , the analysis threshold will be relaxed proportionally, thereby improving the sensitivity to implicit language. This adjustment mechanism ensures that when there are abnormally frequent implicit emotions in public opinion, the system can respond in a timely manner and improve the accuracy and flexibility of public opinion monitoring.

[0045] After the threshold is relaxed, the new analysis threshold is applied to the deep learning model to improve the sensitivity to suggestive and implicit language. At this time, the system enters a verification and feedback stage to verify whether the adjusted analysis threshold effectively improves the ability to capture implicit language signals. Specifically, the system will use the updated analysis threshold to monitor the public opinion topic again and calculate a new round of frequency coefficients . If the new frequency coefficient continues to exceed , indicating that the frequent occurrence of implicit language continues, the system will further strengthen the monitoring intensity and trigger a second round of threshold adjustment to enhance the monitoring depth. Conversely, if the new frequency coefficient does not exceed the adjusted analysis threshold, the system will gradually return to the original analysis threshold setting and enter the normal monitoring state. In this way, the model can adaptively adjust its monitoring strategy to ensure high sensitivity to frequently occurring suggestive and implicit language while avoiding wasting resources on low-frequency public opinion signals. The formula is described as: , If , trigger the second round of threshold adjustment; otherwise, restore the initial threshold .

[0046] By dynamically monitoring the frequency changes of suggestive and veiled language in public opinion topics, potential social risks can be identified in a timely manner. When the frequency of veiled or suggestive language in a certain topic increases, the system automatically triggers a threshold relaxation mechanism to adjust the analysis threshold in real time, enhancing the sensitivity to such implicit signals. This ensures that the model can capture potential emotional changes or social instability factors in a timely manner when faced with complex or indirect expressions, thereby providing more accurate and timely intelligence for the police department and avoiding missing the opportunity to intervene at the initial stage of high-risk public opinion events. This step enables the monitoring system to have stronger flexibility and dynamic response capabilities by adaptively adjusting the analysis threshold, capable of adjusting the monitoring intensity according to the evolution of public opinion to ensure the effective identification and response to potential threats.

[0047] The present invention effectively solves the problem of the decline in the ability to identify subtle public opinion signals due to model lightweighting in the police field through a lightweight large model deployment method applicable to resource-constrained environments. By classifying and deeply analyzing public opinion topics on social media platforms, the model can dynamically monitor topic changes and identify suggestive or veiled language in real time. When the frequency of such language increases, the system can adaptively adjust the analysis threshold to enhance the sensitivity to potential social instability factors, thereby avoiding misjudgment and missing the opportunity to intervene. This solution not only improves the accuracy and efficiency of public opinion monitoring but also ensures that in a resource-constrained environment, the police department can identify and respond to potential social threats in a timely and accurate manner, thus effectively preventing the occurrence of social unrest or violent events at the initial stage.

[0048] The above formulas are all dimensionless and take their numerical values for calculation. The formula is obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation. The preset parameters in the formula are set by those skilled in the art according to the actual situation.

[0049] As mentioned above, only the specific implementation manners of this application are described, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claimed rights.

[0050] Only some exemplary embodiments of the present invention have been described by way of illustration above. Undoubtedly, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

Claims

1. A lightweight large model deployment method applicable to resource-constrained environments, characterized in that, It includes the following steps: Through the API interface or web crawler technology of the social media platform, comprehensively collect and obtain all potential public opinion topics, construct a basic data set for public opinion analysis, and comprehensively grasp the current public opinion dynamics on social media; Classify the collected public opinion data according to topics, and establish an analysis set for each public opinion topic separately to process each public opinion topic independently; During the analysis process, for each public opinion topic, screen out the text data containing suggestive or implicit language, conduct in-depth analysis on the screened data, input the data after in-depth analysis into the pre-trained deep learning model, and evaluate the occurrence frequency of suggestive and implicit language in each public opinion topic; Dynamically monitor the changes of public opinion topics through the deep learning model, and real-time track whether a certain topic frequently appears suggestive or implicit language. When it is found that the frequency of suggestive and implicit language in a certain public opinion topic increases, the threshold relaxation mechanism will be automatically triggered to adaptively adjust the analysis threshold and improve the sensitivity to suggestive and implicit language.

2. The lightweight large model deployment method applicable to resource-constrained environments according to claim 1, wherein The public opinion topics are based on keywords, popular tags, event names or regional divisions, and are used to accurately identify and locate the focus of attention on current social media.

3. The lightweight large model deployment method applicable to resource-constrained environments according to claim 1, wherein, The deep learning model is a natural language processing model based on the Transformer architecture.

4. The lightweight large model deployment method applicable to resource-constrained environments according to claim 1, characterized in that For each public opinion topic, screen out the text data containing suggestive or implicit language through the following steps: Use the sentiment analysis model in natural language processing technology to extract suggestive language expressions by identifying the implicit emotions or feelings in the text; Use keyword matching and semantic analysis to capture the words and phrases with potential suggestive or implicit expressions; Apply the technology of context understanding, and combine the context information to judge whether the public opinion words contain implicit or suggestive content; Combine the sentiment dictionary and semantic role annotation, compare the sentiment tendency of the words with their usage in the context, and further identify the language features containing implicit or suggestive content.

5. The lightweight large model deployment method applicable to resource-constrained environments according to claim 1, characterized in that Conduct in-depth analysis on the screened data, input the data after in-depth analysis into the pre-trained deep learning model, and evaluate the occurrence frequency of suggestive and implicit language in each public opinion topic. The specific steps are as follows: Extract the features reflecting the high-frequency occurrence of suggestive and implicit language from the screened data. Among them, the extracted features include the frequency of fuzzy or ambiguous language expressions in the text data and the frequency of transmitting information in a non-literal way in the text data. After in-depth analysis of the screened data, generate the fuzzy expression ratio reference value and the non-literal communication reference value respectively. Through the fuzzy expression ratio reference value and the non-literal communication reference value, initially quantify the occurrence frequency of suggestive and implicit language in the text data and its potential emotional intensity to help identify potential social instability risks; Input the fuzzy expression ratio reference value and the non-literal communication reference value into the pre-trained deep learning model, generate a frequency coefficient through the model, and evaluate the occurrence frequency of suggestive and implicit language in each public opinion topic based on the frequency coefficient.

6. The lightweight large model deployment method applicable to resource-constrained environments according to claim 5, wherein, The specific steps for generating the fuzzy expression ratio reference value after in-depth analysis of the frequency of fuzzy or ambiguous language expressions in the text data are as follows: In text data, first, ambiguity recognition is performed on all language expressions through natural language processing techniques. For each piece of text data, an ambiguity vocabulary library and a model based on deep semantic parsing are used to identify the ambiguous language expressions therein and assign a weighted score to the ambiguous words. The specific assignment formula is as follows: ， Among them, is the weighted score for the fuzzy expression in the text ; is the fuzziness score of the word w in the fuzziness vocabulary, reflecting the ambiguity degree of the word; is the set of semantic roles of the word w in the context; is the context weight of the semantic role c , calculated based on the deep semantic parsing model, used to reflect the specific meaning and potential emotion of the word in this context; After the ambiguity scoring of the text, the next step is to calculate the reference value of the ambiguity expression ratio. By summarizing the ambiguity scores in all texts, the overall ambiguity expression ratio is obtained, further quantifying the frequency of occurrence of suggestive and implicit language and its potential emotional intensity in this topic. The specific quantification formula is as follows: ， Among them, is the reference value of the fuzzy expression ratio T for the topic; is the set fuzzy expression threshold, used to screen out important fuzzy languages; represents the number of texts that meet the conditions, that is, the number of texts with significant fuzzy expressions.

7. The lightweight large model deployment method applicable to resource-constrained environments according to claim 5, wherein The specific steps for generating the non-literal communication reference value after in-depth analysis of the frequency of information transmission in a non-literal manner in text data are as follows: When analyzing the frequency of information transmission in a non-literal manner in the text, first, an identification model for non-literal communication needs to be established. This identification model is based on natural language processing techniques, combines multiple semantic analysis methods, extracts the characteristics of suggestive, implicit, and indirect language, and assigns a non-literal communication weight to the non-literal expression. The calculation expression of the non-literal communication weight is as follows: ， Among them, represents the non-literal communication weight in text t, is the non-literal expression intensity function of text t in the context change, is the weight parameter of each non-literal language feature, is the total length or total time period of the text; After obtaining the non-literal weight of each text, the obtained non-literal weights are normalized and combined to generate the non-literal communication reference value. The generated expression is as follows: ， Among them, is the reference value for non-literal communication, which is the adjustment factor for each text segment.

8. The lightweight large model deployment method applicable to resource-constrained environments according to claim 5, wherein When it is found that the frequency of suggestive and implicit language in a certain public opinion topic increases, the threshold relaxation mechanism is automatically triggered to adaptively adjust the analysis threshold. The specific steps are as follows: When the frequency coefficient exceeds the frequency coefficient reference threshold the threshold relaxation mechanism is automatically activated to adjust the analysis threshold. The new analysis threshold is adjusted by the following formula, and the adjustment expression is: ， Among them, is the current analysis threshold, is the adjusted analysis threshold; After the threshold is relaxed, the new analysis threshold is applied to the deep learning model and follows the following formula: ， If , trigger the second round of threshold adjustment; otherwise, restore the initial threshold .

Images