A network catchword automatic identification and contextualized explanation generation method

Abstract

The application discloses a network popular language automatic identification and contextual explanation generation method. The method comprises the following steps: constructing and dynamically maintaining a network popular language knowledge base, and performing real-time active identification on the network popular language in the social media browsing content based on a search enhancement generation technology; modeling the explicit features of a user, obtaining demographic information and value preference information; for the identified network popular language, generating an accurate definition based on a user-generated content (UGC) driven multi-dimensional semantic induction, and through a contextual explanation generation method constrained by a semantic framework, converting the abstract definition into a micro story fitting the personal experience of the user; and realizing the end-side light weighting of the explanation model through step-by-step knowledge distillation and low-bit quantization technology. The application can provide low-interference, active network popular language identification and hierarchical explanation services for users in a social media browsing scene, and effectively reduce the language understanding load in digital interaction.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and human-computer interaction technology, specifically, it relates to a method for automatic identification and contextualized interpretation of popular internet slang. Background Technology

[0002] With the popularization of mobile internet technology and the widespread use of online social media, internet slang has become an important carrier of daily digital communication. Internet slang is usually generated during online interactions and spreads rapidly through social media. Its word formation is diverse, including abbreviations, homophones, neologisms, and new meanings of old words. However, internet slang is characterized by its semantic dependence on context, diverse structures, and rapid updates, making its meaning often difficult to infer directly from literal information, easily leading to misunderstandings. Incomprehensible internet slang increases the burden on users browsing social media content and may even lead to misjudgments of attitudes and emotional orientations towards existing content, causing social isolation and digital rejection. Therefore, there is an urgent need for a method and system that can proactively identify internet slang during social media browsing and provide users with accurate, contextualized, and personalized explanations. The following sections introduce the current research status and shortcomings of internet slang identification and explanation generation technologies.

[0003] (1) Internet slang recognition technology

[0004] Automatic identification of internet slang is a prerequisite for generating explanations. Existing methods for identifying internet slang mainly fall into two categories: dictionary-based methods and machine learning-based methods. Dictionary-based methods rely on a pre-built lexicon of internet slang terms, detecting slang terms from text through string matching or regular expressions. These methods are simple to implement and fast, but face two limitations: first, internet slang updates extremely quickly, and static dictionaries struggle to cover constantly emerging new words and new meanings of old words; second, the same expression may be both a common word and an internet slang term in different contexts (e.g., "Versailles" can refer to a French place name or a subtle way of showing off), making it easy to misjudge or miss the correct term if relying solely on rule matching. Machine learning-based methods model internet slang identification as a sequence labeling or text classification task, using models such as Conditional Random Fields and Bidirectional Long Short-Term Memory Networks to learn recognition capabilities from labeled corpora. In recent years, the emergence of pre-trained language models such as BERT and RoBERTa has significantly improved the accuracy of identification.

[0005] The aforementioned methods achieve some success in scenarios with fixed word lists or high-frequency terms, but they still have significant limitations in the real-time browsing environment of social media. On the one hand, internet slang updates rapidly and has many variations, making it easy to miss or lag when relying solely on static dictionaries or rule matching. On the other hand, for expressions with new meanings of old words, strong contextual dependence, or cross-sentence triggering, relying solely on literal matching is insufficient to accurately determine whether a term in the current context constitutes an internet slang term, thus affecting the accuracy of subsequent interpretations. Furthermore, if the recognition module directly serves the user's social media browsing process, it also needs to consider latency, stability, and low interference, which existing methods do not adequately support.

[0006] (2) Internet slang explanation generation technology

[0007] The technology for generating explanations of popular online slang aims to produce natural language definitions for target slang terms, helping users understand their semantic and pragmatic functions. Existing technological approaches mainly include structured explanations based on knowledge bases, crowdsourced explanations based on user-generated content (UGC), and automated explanations based on generative models. Knowledge base-based methods support explanation generation by constructing a knowledge graph containing definitions, synonyms, usage examples, and socio-cultural background information for the slang term. The advantage of this approach lies in the authority and consistency of the explanation content, but its construction and maintenance costs are high, and the speed of knowledge updates cannot keep up with the rapid evolution of slang terms. Crowdsourced explanations based on UGC primarily take the form of online dictionaries (such as Urban Dictionary), gathering explanation resources through user-submitted entries and definitions. This approach can quickly respond to emerging slang terms, but the quality of content varies greatly, lacking unified definition standards and review mechanisms, making it difficult to guarantee the accuracy and authority of the explanations. For users, manually searching the crowdsourced dictionary requires switching between the source social media interface and the dictionary interface, which is a heavy operational burden. In addition, the explanations are mainly general definitions, which are lengthy and complex in language, and lack adaptation design for users' comprehension ability and cognitive characteristics.

[0008] Automated interpretation based on generative models has become a research hotspot in recent years. Early work formalized the definition of the generative task as a mapping problem from a continuous vector space to discrete natural language sequences, generating interpretations by training neural network models given target words and their context. Noraset et al. first proposed defining a modeling task, the core objective of which can be formalized as maximizing conditional probabilities:

[0009]

[0010] in, Target popular phrases, The generated definition sequence is shown in the figure. Subsequently, Gadetsky et al. proposed a context-aware definition generation framework, which captures cue words in the input sentence by introducing an attention mechanism. With the advent of pre-trained language models, the model can more fully utilize dynamic contextual representations to model polysemy, providing more suitable input signals for generating definitions with different meanings of the same word.

[0011] However, directly applying existing generative methods to the explanation of internet slang still faces three key challenges. First, definition resources are outdated: internet slang is a typical example of out-of-vocabulary words. Mainstream pre-trained models are trained on static corpora, making it difficult to learn the rapidly changing semantics of emerging slang in a timely manner, resulting in inadequate inference and semantic drift in the generated explanations. Second, contextualization and personalization are lacking: most existing methods output standardized dictionary-style definitions, lacking the ability to map abstract definitions to specific usage contexts, and failing to consider the impact of users' personal background and value preferences on the acceptance of explanations. Research shows that users tend to accept contextualized explanations that are close to their own life experiences, and are prone to aversion when the implicit value orientation in the explanation is inconsistent with their concepts of "stability, respect, and harmony." Third, on-device deployment efficiency is low: while large language models possess powerful generation capabilities, their large parameter scale and high inference latency make it difficult to achieve real-time explanation services in multi-user concurrent scenarios on mobile devices. How to achieve model lightweighting while maintaining explanation quality is a significant bottleneck restricting the practical application of the technology. In summary, existing technologies still have significant shortcomings in the active identification of internet slang, the generation of contextualized personalized explanations, and the lightweighting of models, and a systematic solution is urgently needed. Summary of the Invention

[0012] To address the shortcomings of existing technologies, the present invention aims to provide a method for automatically identifying and generating contextualized explanations of internet slang, thereby solving the technical problems of inaccurate identification of internet slang, lagging definition resource updates, lack of personalization and contextual adaptability of explanations, and difficulty in deploying large-scale models on edge devices. This invention can provide users with low-interference, proactive internet slang identification and hierarchical explanation services in social media browsing scenarios, effectively reducing the language understanding load in digital interactions.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] This invention provides a method for automatically identifying and generating contextualized interpretations of popular internet slang, comprising the following steps:

[0015] Step 1: Automatic Identification of Internet Buzzwords: Construct and maintain a knowledge base of internet buzzwords. Use the target text as a query to retrieve relevant entries from the knowledge base as an evidence set. Input the target text and the evidence set into a large language model. The large language model will judge based on the context and output a candidate set of internet buzzwords. When the evidence retrieved from the knowledge base is insufficient, guide the large language model to make supplementary judgments based on its own capabilities or online searches, and expand the knowledge base with newly identified internet buzzwords and their definitions through an incremental update mechanism.

[0016] Step 2, User Feature Modeling: Collect users' demographic information and value preference information, and transform the demographic information and value preference information into structured user features, which will serve as input for personalized conditions generated in subsequent interpretations;

[0017] Step 3: Generation of Internet Buzzword Definitions Based on User-Generated Content: For the target internet buzzwords identified in Step 1 and their corresponding user-generated content sets, a three-stage automated screening mechanism is used to select a high-quality subset of evidence. Combined with multi-dimensional semantic induction, an accurate definition of the target internet buzzword is generated. The three-stage automated screening mechanism includes a coarse screening of candidate user-generated content based on a weighted average word uncertainty score model, a fine screening of semantic consistency based on anti-factual entity perturbation, and a deduplication screening based on density clustering.

[0018] Step 4: Contextualized Explanation Generation Based on Semantic Framework Constraints: Construct a semantic framework based on the definitions generated in Step 3 and determine the core framework elements; fill in the framework elements according to the user characteristics generated in Step 2 and the high-quality evidence subset; generate narrative style control signals to constrain the tone and value orientation of the story; and output a contextualized explanation in the form of a micro-story by first generating a story outline and then generating a complete story.

[0019] In this invention, in step one, the internet slang knowledge base uses internet slang terms as the primary key and maintains a basic definition for each term. The specific process of the incremental update mechanism is as follows: when the large language model identifies a new term not covered by the knowledge base in the target text, the new term is regarded as a candidate incremental knowledge, and a basic definition is generated according to the internet slang definition generation method based on user-generated content described in step three. Then, the new term and the basic definition are written back to the knowledge base to complete the expansion.

[0020] In this invention, in step two, the demographic information includes age, gender, occupation, and education level; the value preference information is collected through the Schwarz Portrait Values ​​Questionnaire-21, and the user's score on the basic value dimension is calculated; the structured user characteristics are stored in JSON format.

[0021] In this invention, the specific process of the three-stage automated screening mechanism in step three is as follows:

[0022] Phase 1: Coarse screening of candidate user-generated content based on a weighted average word uncertainty score model, followed by masking strategy to further refine candidate use cases (i.e., the candidate set of user-generated content). Target Internet slang terms Replace with the special tag [MASK], and obtain the context embedding vector of the [CLS] tag through BERT model encoding. The initial correlation score S0 was obtained by multilayer perceptron (MLP) scoring. , )=σ(MLP( )), select Top-M candidates, where σ is the sigmoid activation function;

[0023] Phase Two: Semantic Consistency Refinement Screening Based on Counterfactual Entity Perturbation. Three perturbation strategies are employed: replacing target internet buzzwords with [MASK], replacing them with random words, and replacing other entities in the sentence with [MASK]. Consistency scores for direct interpretation and entity substitution interpretation are calculated respectively, yielding a semantic consistency score μ and a robustness score r. The final refinement score is:

[0024]

[0025] Where α, β, and λ are adjustable parameters, and penalty is a conflict penalty term;

[0026] Phase 3: Deduplication screening based on DBSCAN density clustering. User-generated content is embedded with vector encoding and density clustering. The highest-scoring sample is retained from each cluster to obtain a high-quality evidence subset.

[0027] In this invention, step three involves the following steps: For each piece of user-generated content in the high-quality evidence subset, a six-dimensional definition is generated using a large language model. This six-dimensional definition includes intent, concept, grammar, social cue, context, and spelling / pronunciation dimensions. These six-dimensional definitions are then integrated into a complete definition of the target online buzzword through an inductive fusion step. Simultaneously, an iterative self-evaluation mechanism is introduced to calculate the support degree of the user-generated content for the current version definition. The new sorting score is formed by merging the fine screening score S1 with the new score:

[0028]

[0029] Where γ is the support weight and t is the iteration round.

[0030] In this invention, in step four, the semantic framework is constructed based on the theory of frame semantics, including the frame name, frame definition, and core frame elements; the core frame elements are used to describe the indispensable participants and key relationships in the context, and to distinguish between core elements and peripheral elements. The core elements are key information that determines the frame recognition category, and the peripheral elements are optional supplementary contextual information.

[0031] In this invention, in step four, the narrative style control signal is generated as follows: the ten dimensions of Schwartz's values ​​are mapped to a four-dimensional space, which includes openness, conservatism, self-improvement, and self-transcendence; where openness corresponds to the self-direction and stimulation dimension, conservatism corresponds to the tradition and conformity dimension, self-improvement corresponds to the achievement and power dimension, and self-transcendence corresponds to the philanthropy dimension; the narrative stance is determined based on the user's score distribution on the four dimensions; the offense avoidance intensity is set based on the user's safety dimension score; the strategy of first generating a story outline and then generating a complete story specifically includes: generating a sentence-level story outline in the planning stage, explicitly allocating the core framework elements to 3 to 5 sentences of narrative structure, clearly specifying the appearance position of the target internet buzzword in the story, so that the target internet buzzword appears exactly once in the story; and generating the story text sentence by sentence under the constraints of the story outline in the generation stage, ensuring that the coverage rate of the core framework elements reaches a set threshold.

[0032] This invention also includes step five:

[0033] Explanation Model Lightweighting: A stepwise knowledge distillation method is used to compress a large-scale language model into a lightweight student model, and the output quality is improved by combining preference optimization methods. Low bit weight quantization and inference memory management reduce memory usage and latency, making it easier to deploy on the edge.

[0034] In this invention, step five, stepwise knowledge distillation, includes:

[0035] The training dataset is constructed, and intermediate results of each step, including semantic framework construction, personalized filtering of user-generated content, filling of core framework elements, narrative style control, story outline generation, and final story generation, are added to the original data.

[0036] Step-by-step structured supervision trains student models in the order of semantic framework generation and core framework element tasks, story outline and story generation tasks, enabling the models to gradually learn narrative style and personalization capabilities based on their mastery of semantic element generation capabilities.

[0037] In this invention, in step five, the preference optimization method adopts the ORPO method, and the low-bit weight quantization adopts 4-bit integer quantization. The quantization formula is:

[0038]

[0039] in, S is the original 32-bit floating-point weight, S is the scaling factor, and Z is the zero-point offset.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] 1. This invention employs a retrieval-enhanced generation method for automatic identification of internet slang, combining knowledge base retrieval with large language model discrimination. This ensures both interpretability of the identification and the ability to understand the context, significantly improving the accuracy of internet slang identification. Furthermore, the incremental update mechanism gradually expands the knowledge base, alleviating the cold start problem of the knowledge base.

[0042] 2. This invention proposes a three-stage automated screening mechanism. Through the synergistic effect of three stages—coarse screening using a weighted average word uncertainty score model, fine screening using counterfactual perturbation, and deduplication using density clustering—it effectively suppresses the interference of noisy corpora on definition induction. Combined with an iterative self-evaluation mechanism, it further improves the reliability of definition generation.

[0043] 3. This invention proposes a contextualized explanation generation method based on semantic frame constraints. It organizes the story structure with frame semantics theory and ensures the structural and semantic consistency of the generated content through a narrative strategy of planning before generation. It introduces user feature modeling and narrative style control so that the generated explanation can match the user's value orientation and understanding level, thus realizing personalized contextualized explanation.

[0044] 4. This invention adopts a hierarchical interpretation output mechanism, which provides users with a progressive understanding path from rapid semantic mapping to deep contextual understanding through a combination of personalized metaphor interpretation and personalized micro-story interpretation.

[0045] 5. This invention uses lightweight techniques such as stepwise knowledge distillation, ORPO preference optimization, and low-bit quantization to compress large-scale language models into small-scale models that can run efficiently on edge devices, providing support for edge deployment and real-time response of popular internet slang explanation services. Attached Figure Description

[0046] Figure 1 This is a system architecture diagram of the method for identifying and generating contextualized interpretations of popular online slang in this invention.

[0047] Figure 2 This is a system prototype diagram of the method for identifying and generating contextualized interpretations of popular online slang in this invention.

[0048] Figure 3 This is a flowchart of the method for generating contextualized explanations of popular online slang according to the present invention. Detailed Implementation

[0049] This invention provides a method for identifying and generating contextualized interpretations of popular internet slang, the overall system framework of which is as follows: Figure 1 As shown, the method prototype diagram is as follows: Figure 2 As shown. This method consists of two stages: automatic identification of internet slang and contextualized interpretation generation. The first stage includes three main steps, and the second stage includes four main steps, as detailed below.

[0050] I. The Stage of Identifying Popular Internet Slang Based on Dynamic Knowledge Base

[0051] Figure 1 The slang recognition module in the example demonstrates the workflow of this step. This module is designed for users browsing short videos on social media. Its core task is to automatically identify slang terms in the video text information that the user is reading or watching, and output a candidate set of slang terms to be explained.

[0052] Step 1: Building a Knowledge Base for Popular Internet Slang

[0053] Construct a knowledge base with popular internet slang terms as the primary key, and maintain a basic definition for each term. Terms are used to support evidence recall during the retrieval phase, while basic definitions provide stable definition input for subsequent contextualized interpretation generation. The knowledge base structure can be formally represented as:

[0054]

[0055] in Indicates the first A popular internet slang term. This indicates its corresponding basic definition.

[0056] Step 2: Retrieval Enhancement Recognition

[0057] For input text The system first performs preprocessing and normalization operations, including normalizing differences in capitalization, whitespace, and other formalities. The normalized text is then used as the query in the knowledge base. The most relevant set of entries retrieved from the database is denoted as the evidence set. This evidence set consists of several terms and their basic definitions, providing the model with comparable candidate clues and definitional evidence.

[0058] The input text and evidence set are provided to the large language model, which then judges candidate terms in the context and outputs a set of popular internet slang. The recognition process can be formalized as follows:

[0059]

[0060] in This represents the inference function of a large language model. The model needs to explain its judgment criteria based on traceable data fragments to improve the interpretability of recognition.

[0061] Step 3: Incremental update of the knowledge base

[0062] When the model identifies new terms not covered by the knowledge base At this point, the system treats the term as a candidate incremental knowledge and generates its basic definition according to the definition generation method described in step three. Then, the terms and definitions are written back to the knowledge base to complete the expansion:

[0063]

[0064] When the retrieval phase fails to retrieve sufficient evidence from the knowledge base, i.e. When the text is empty or lacks support for the current context, the system guides the model to judge the text based on its own language understanding capabilities. If further verification is needed, it calls the online retrieval function to obtain supplementary evidence, and then combines the text context to make a final judgment on the candidate terms. Through this incremental maintenance mechanism driven by real needs, the system can gradually expand the scope of explanation while ensuring the stability of interaction, so that the recognition capability continues to improve with use.

[0065] II. The Contextualized Explanation and Generation Stage of Popular Internet Slang Based on User-Generated Content and Semantic Framework Constraints

[0066] The core task of this phase is to model user characteristics and generate two levels of explanations for the popular internet slang identified in the first phase: the first level is the explanation of metaphorical terms, and the second level is the explanation of contextualized stories. Simultaneously, to achieve real-time service on mobile devices, the explanation model is made lightweight.

[0067] Step 1: User Feature Modeling

[0068] User characteristic modeling explicitly models users from two dimensions: demographic information and value information, providing structured conditional inputs for subsequent contextualized interpretation generation.

[0069] (1) Collection of population statistics information

[0070] The system collects basic information about users, such as age, gender, education level, occupation, and hobbies, and has users fill out a questionnaire.

[0071] (2) Collection of value information

[0072] The Schwarz Portrait Values ​​Questionnaire (PVQ-21) was used for measurement. The questionnaire contains 21 questions on the person profile, covering ten basic value dimensions: self-orientation, power, achievement, security, stimulation, conformity, tradition, hedonism, and kindness. A 6-point rating scale was used (1 = "very much like me" to 6 = "very unlike me").

[0073] (3) Value score calculation

[0074] The questionnaire items were arranged in reverse order so that higher scores represented greater value. Value scores for users across ten value dimensions were calculated following European social survey guidelines. (The last sentence appears to be incomplete and possibly refers to a user's score.) The original score on the i-th dimension is Then the value fraction is:

[0075]

[0076] Further centralization is used to correct for individual differences:

[0077]

[0078] in For users The average score across all ten dimensions.

[0079] (4) Structured user feature output

[0080] Demographic information and value scores are converted into a fixed data structure format and stored in JSON format. This structured user profile will serve as input for interpreting the generated personalized conditions.

[0081] Step 2: Generating Popular Internet Slang Definitions Based on User-Generated Content

[0082] Since internet slang is a typical out-of-vocabulary word, its semantics are highly dependent on context and updated rapidly. This invention adopts a definition generation method based on user-generated content (UGC), which generates accurate definitions through three-stage automated screening and multi-dimensional semantic induction. Figure 3 It demonstrates the complete definition generation process.

[0083] (1) First stage coarse screening - candidate UGC coarse screening based on WAUS scoring

[0084] A semantically independent use case selector model is used to score the quality of candidate UGC. This model replaces target buzzwords with masked symbols using a masking strategy, forcing the model to judge the quality of UGC content based on contextual clues to determine whether the UGC content possesses interpretable evidence. Given a buzzword... With candidate use cases First, construct the masked version. ,Will The target buzzword in the middle is replaced with the corresponding token. Then, the BERT encoder was used to... Encoding yields vector representation Output quality scores via a lightweight MLP classification head:

[0085]

[0086] in Using the sigmoid activation function, select the top-scoring results in descending order of scores. Items as a preliminary screening set of evidence .

[0087] (2) Second-stage fine screening - UGC fine screening based on semantic consistency and counterfactual perturbation

[0088] In social media user-generated content (UGC), single interpretations are easily influenced by external knowledge, trending events, emotional expressions, and other entity factors, leading to seemingly reasonable but inaccurate explanations from the model. This invention constructs counterfactual contexts through entity substitution and examines the consistency of interpretations under perturbations.

[0089] Direct interpretation of each candidate UGC construction Explanation of entity substitution Consistency of computation definition and robustness score :

[0090]

[0091]

[0092] in This indicates the degree of consistency defined under counterfactual perturbations. This indicates instability. Furthermore, we define a robustness score:

[0093]

[0094] The final evidence quality score is defined as:

[0095]

[0096] in , The relative importance of control consistency and robustness Used to penalize models for situations where content is blocked or the meaning is ambiguous. The function is an indicator function; it takes a value of 1 if the condition is true, and 0 otherwise. Based on this score constraint, high-quality evidence should maintain semantic stability under local changes in context, thus being more likely to capture the core semantics of internet slang rather than local details. After scoring, for each slang term... , sort the candidate set by Sort the evidence in descending order and select the top K pieces as the set of fine-tuned evidence. .

[0097] (3) The third stage of final screening - semantic deduplication based on density clustering

[0098] Calculate the embedding vector for each UGC. Then, DBSCAN clustering is used in the embedding space to treat high-density regions as repeating clusters, resulting in a cluster set. Subsequently, the sample with the highest evidence score is retained within each cluster, resulting in...

[0099]

[0100] Finally, the evidence is merged and sorted by score to obtain the final set of evidence for screening. .

[0101] (4) Multidimensional definition and attribution generation

[0102] Selecting high-quality UGC collections Then, the large language model is guided to generate definition descriptions from six dimensions: intent understanding, concept formation, grammatical understanding, social cues, context, and spelling and pronunciation. The model needs to explain the reasoning process before giving the definitions for each dimension.

[0103] (5) Definition, induction and iterative self-evaluation

[0104] The six-dimensional definitions are integrated and summarized into a complete definition. An iterative self-evaluation mechanism is introduced, utilizing the already generated definitions. Reverse-evaluate the support of each UGC for the definition:

[0105]

[0106] Combined with the evidence quality score, a new ranking score is formed:

[0107]

[0108] Step 3: Contextualized Explanation Generation Based on Semantic Frame Constraints

[0109] Figure 3 This paper demonstrates the complete workflow of a contextualized explanation generation method based on semantic framework constraints. After obtaining the definitions of popular internet slang and user characteristic information, the abstract definitions are transformed into understandable contextualized descriptions.

[0110] First, based on frame semantics theory, the definition of internet slang is used as a boundary constraint for frame semantics, and high-quality UGC collections are included. As a source of evidence for framework elements; treating core framework elements as semantic slots, in The system selects content similar to the user's characteristics for filling in the blanks; it then calculates a four-quadrant value score based on the user's ten-dimensional basic value score, which is mapped to a narrative style control signal. The four quadrants include conservative (polite and low-conflict), open (encouraging understanding of new things), self-transcendence (emphasizing consideration and care), and self-improvement (emphasizing efficiency and recognition). Openness corresponds to the self-directed and stimulation dimension, conservative corresponds to the tradition and conformity dimension, self-improvement corresponds to the achievement and power dimension, and self-transcendence corresponds to the philanthropic dimension.

[0111] All narratives are constrained to be polite and avoid boasting or belittling others; finally, a two-stage strategy of planning first and then generating is adopted. The first stage generates a story outline, explicitly assigning core framework elements to the narrative structure; the second stage generates the story text under the constraints of the outline and outputs alignment information.

[0112] Step 4: Explain the lightweighting of the model

[0113] To achieve real-time explanation services in multi-user scenarios on mobile devices, this invention adopts the following scheme for model lightweighting.

[0114] (1) Construction of training dataset

[0115] The model incorporates intermediate results from each step of the original data process: semantic framework construction, UGC personalization filtering, core framework element filling, narrative style control, story outline generation, and final story generation. Each sample consists of a condition section and a target section.

[0116] (2) Stepwise knowledge distillation

[0117] Low-rank supervised fine-tuning is employed, with two training phases. The first phase involves learning semantic frame generation and core frame elements.

[0118]

[0119] The second phase involves learning the story outline and story generation tasks:

[0120]

[0121] (3) Preference optimization

[0122] The ORPO method is used for preference alignment. The overall quality score is defined as:

[0123]

[0124] The objective function for ORPO optimization is:

[0125]

[0126] in For the current strategy model, As a reference model, Output according to preference. This is the output for non-preferred users.

[0127] (4) Weight compression and inference memory management

[0128] Low-bit weight quantization is used, prioritizing the quantization of the main model weights:

[0129]

[0130] To comprehensively evaluate the performance of the proposed method on the contextualized explanation generation task, this paper conducts a systematic experimental comparison on the Chinese Internet Buzzword Explanation with Evidence and Rationales dataset CHEER. The CHEER dataset contains 1126 popular internet slang terms, each accompanied by a set of user-generated content (UGC) and a manually written reference definition. Evaluation models include the baseline method (directly calling a large language model to generate explanations) and models of different sizes (Qwen2.5-0.5B, 1.5B, 3B, 7B, and Qwen-Max) guided by the proposed method (OURS). Evaluation metrics cover four dimensions: explanation quality, story quality, intermediate result quality, and hard constraints. Experimental results are shown in Tables 1 and 2.

[0131] Table 1. Performance Comparison of Generative Models

[0132] As shown in Table 1, the method of the present invention significantly outperforms the baseline method in both semantic accuracy (SA) and detail integrity (SC).

[0133] Table 2. Performance comparison of the models on the contextualized interpretation generation task.

[0134] As shown in Table 2, the method of this invention has four main findings: 1. The method of this invention significantly improves semantic fidelity and structural controllability. Qwen-Max achieves the best results in three key indicators: NLI-Contr (contradiction degree), FEC (core element coverage), and TWC (target word hit rate), proving the effectiveness of semantic framework constraints and the pre-planning and post-generation strategy.

[0135] 2. Model size has a significant impact on constraint compliance. Although the 0.5B model performs well in terms of diversity and partial condition generation, it is far inferior to the large model in FEC (0.2519) and TWC (0.4275), indicating that the very small model has difficulty understanding and executing complex structured generation instructions.

[0136] 3. The distillation models each have their advantages in terms of fluency and personalized filling. The 1.5B model has the lowest PPL (14.09), while the 7B model has the highest FEF (0.8079), indicating that the lightweight models are competitive in specific sub-tasks, but Qwen-Max still has the most balanced overall performance.

[0137] 4. Personalized explanation generation requires a multi-dimensional balance. Simply pursuing fluency or diversity cannot guarantee the effectiveness of the explanation. This invention achieves the best balance between semantic accuracy, format controllability, and personalized expression through structured constraints, providing users with an explanation service for popular internet slang that is easy to understand and does not deviate from the original meaning.

Claims

1. A method for automatically identifying and contextualizing popular internet slang, characterized in that, Includes the following steps: Step 1: Automatic Identification of Internet Buzzwords: Construct and maintain a knowledge base of internet buzzwords. Use the target text as a query to retrieve relevant entries from the knowledge base as an evidence set. Input the target text and the evidence set into a large language model. The large language model will judge based on the context and output a candidate set of internet buzzwords. When the evidence retrieved from the knowledge base is insufficient, guide the large language model to make supplementary judgments based on its own capabilities or online searches, and expand the knowledge base with newly identified internet buzzwords and their definitions through an incremental update mechanism. Step 2, User Feature Modeling: Collect users' demographic information and value preference information, and transform the demographic information and value preference information into structured user features, which will serve as input for personalized conditions generated in subsequent interpretations; Step 3: Generation of Internet Buzzword Definitions Based on User-Generated Content: For the target internet buzzwords identified in Step 1 and their corresponding user-generated content sets, a three-stage automated screening mechanism is used to select a high-quality subset of evidence. Combined with multi-dimensional semantic induction, an accurate definition of the target internet buzzword is generated. The three-stage automated screening mechanism includes a coarse screening of candidate user-generated content based on a weighted average word uncertainty score model, a fine screening of semantic consistency based on anti-factual entity perturbation, and a deduplication screening based on density clustering. Step 4: Contextualized Explanation Generation Based on Semantic Framework Constraints: Construct a semantic framework based on the definitions generated in Step 3 and determine the core framework elements; fill in the framework elements according to the user characteristics generated in Step 2 and the high-quality evidence subset; generate narrative style control signals to constrain the tone and value orientation of the story; and output a contextualized explanation in the form of a micro-story by first generating a story outline and then generating a complete story.

2. The method according to claim 1, characterized in that, In step one, the internet slang knowledge base uses internet slang terms as the primary key and maintains a basic definition for each term. The specific process of the incremental update mechanism is as follows: when the large language model identifies a new term not covered by the knowledge base in the target text, the new term is regarded as a candidate incremental knowledge, and a basic definition is generated according to the user-generated content-based method described in step three. Then, the new term and the basic definition are written back to the knowledge base to complete the expansion.

3. The method according to claim 1, characterized in that, In step two, the demographic information includes age, gender, occupation, and education level; the value preference information is collected through the Schwarz Portrait Values ​​Questionnaire-21 to calculate the user's score on the basic value dimension; and the structured user characteristics are stored in JSON format.

4. The method according to claim 1, characterized in that, In step three, the specific process of the three-stage automated screening mechanism is as follows: Phase 1: Coarse screening of candidate user-generated content based on a weighted average word uncertainty score model, followed by masking strategy to filter candidate use cases. Target Internet slang terms Replace with the special tag [MASK], and obtain the context embedding vector of the [CLS] tag through BERT model encoding. The initial correlation score S0 was obtained by multilayer perceptron (MLP) scoring. , )=σ(MLP( We select Top-M candidates, where σ is the sigmoid activation function. Phase Two: Semantic Consistency Refinement Screening Based on Counterfactual Entity Perturbation. Three perturbation strategies are employed: replacing target internet buzzwords with [MASK], replacing them with random words, and replacing other entities in the sentence with [MASK]. Consistency scores for direct interpretation and entity substitution interpretation are calculated respectively, yielding a semantic consistency score μ and a robustness score r. The final refinement score is: ； Where α, β, and λ are adjustable parameters, and penalty is a conflict penalty term; Phase 3: Deduplication screening based on DBSCAN density clustering. User-generated content is embedded with vector encoding and density clustering. The highest-scoring sample is retained from each cluster to obtain a high-quality evidence subset.

5. The method according to claim 1, characterized in that, In step three, the multidimensional semantic induction is as follows: for each piece of user-generated content in the high-quality evidence subset, a six-dimensional definition is generated using a large language model. The six dimensions include intent dimension, concept dimension, grammatical dimension, social cue dimension, context dimension, and spelling / pronunciation dimension. The six-dimensional definition is then integrated into a complete definition of the target internet slang through an inductive fusion step. Simultaneously, an iterative self-evaluation mechanism is introduced to calculate the degree of support for user-generated content to the current version definition. The new sorting score is formed by merging the fine screening score S1 with the new score: ； Where γ is the support weight and t is the iteration round.

6. The method according to claim 1, characterized in that, In step four, the semantic framework is constructed based on the theory of frame semantics, including the frame name, frame definition and core frame elements; the core frame elements are used to describe the indispensable participants and key relationships in the context, and to distinguish between core elements and peripheral elements. The core elements are key information that determines the frame recognition category, and the peripheral elements are optional supplementary contextual information.

7. The method according to claim 1, characterized in that, In step four, the narrative style control signal is generated by mapping the ten dimensions of Schwartz's values ​​to a four-dimensional space, which includes openness, conservatism, self-improvement, and self-transcendence; and determining the narrative stance based on the user's score distribution in the four dimensions. Set the offense avoidance level based on the user's security dimension score; The strategy of first generating a story outline and then generating a complete story specifically includes: During the planning phase, a sentence-level story outline is generated, explicitly assigning core framework elements to 3 to 5 sentences of narrative structure, and clearly specifying the appearance position of the target internet buzzword in the story, so that the target internet buzzword appears exactly once in the story; during the generation phase, the story text is generated sentence by sentence under the constraints of the story outline, ensuring that the coverage rate of the core framework elements reaches the set threshold.

8. The method according to claim 1, characterized in that, It also includes step five: Explanation Model Lightweighting: A stepwise knowledge distillation method is used to compress a large-scale language model into a lightweight student model, and the output quality is improved by combining preference optimization methods. Low bit weight quantization and inference memory management reduce memory usage and latency, making it easier to deploy on the edge.

9. The method according to claim 8, characterized in that, Step five, stepwise knowledge distillation, includes: The training dataset is constructed, and intermediate results of each step, including semantic framework construction, personalized filtering of user-generated content, filling of core framework elements, narrative style control, story outline generation, and final story generation, are added to the original data. Step-by-step structured supervision trains student models in the order of semantic framework generation and core framework element tasks, story outline and story generation tasks, enabling the models to gradually learn narrative style and personalization capabilities based on their mastery of semantic element generation capabilities.

10. The method according to claim 8, characterized in that, In step five, the preference optimization method adopts the ORPO method, and the low-bit weight quantization adopts 4-bit integer quantization. The quantization formula is as follows: ； in, S is the original 32-bit floating-point weight, S is the scaling factor, and Z is the zero-point offset.

Images