Intelligent content filtering method based on specified theme scene

Abstract

The invention relates to an intelligent content filtering method based on a specified theme scene. Accurate filtering of irrelevant content on an ASR output text in the specified theme scene is achieved. The method comprises the following steps: optimizing a zero sample classification model to obtain an optimized zero sample classification model; performing subject classification on the first ASR output text by adopting an optimized zero sample classification model to obtain a first irrelevant content text; constructing a dynamic word bank, calculating the correlation between the first ASR output text and the dynamic word bank, and screening according to the correlation to obtain a second irrelevant content text; performing repeated content comparison on the first irrelevant content text and the second irrelevant content text to obtain a third irrelevant content text; and removing the third irrelevant content text from the first ASR output text to obtain the filtered first ASR output text.

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, and in particular to a method for intelligent content filtering based on a specified theme scenario. Background Technology

[0002] With the rapid development of speech recognition and voice interaction technologies, speech analysis has been widely applied in fields such as information processing, human-computer interaction, intelligent customer service, and speech recognition. However, in practical applications, speech data often contains a large amount of invalid information, such as background noise, irrelevant speech, repetitive content, and invalid commands. This information not only affects the recognition accuracy but also increases the system's processing burden.

[0003] To improve the efficiency of speech information processing, existing technologies often employ various techniques to filter and identify invalid information. For example, speech recognition models based on Natural Language Processing (NLP) and deep learning can perform semantic analysis and classification of speech content, identifying and filtering invalid content. Furthermore, Voice Activity Detection (VAD) technology can be used to identify valid speech segments within speech fragments, filtering background noise and silent segments. Simultaneously, multimodal fusion technologies (such as speech, text, and images) are also used to enhance the accuracy and robustness of information processing.

[0004] However, existing technologies still have many shortcomings in complex scenarios. For example, speech recognition and accuracy remain challenging in complex environments (such as noisy environments, multiple speakers, and far-field speech); some models have limited ability to understand context and struggle to effectively identify invalid information in the context; in addition, model training and deployment costs are high, and they are highly dependent on computing resources and data.

[0005] A search of existing technical literature revealed a patent application (application number 202410727979.6) entitled "A Method and System for Recommending Telecommunications Customer Service Based on Intelligent Voice Analysis." This patent collects customer voice data from a telecommunications customer service system, converts the voice data into text information, and simultaneously acquires the intonation information of the voice data. It then imports the acquired intonation information and historical business urgency information into a business urgency value analysis strategy for analysis. Next, it substitutes the acquired converted text information into a keyword extraction strategy for keyword extraction. Finally, it imports the acquired keywords and historical keywords into a business importance analysis strategy for business importance analysis. Finally, it imports the analyzed business urgency value and business importance into a business recommendation coefficient analysis strategy to obtain a business recommendation coefficient, and makes recommendations to business departments based on the obtained business recommendation coefficient. However, this patent lacks topic classification of the recorded text and uses a pre-set traditional static lexicon, resulting in deficiencies in recognition accuracy, efficiency, and robustness. Summary of the Invention

[0006] Therefore, it is necessary to provide a method for intelligent content filtering based on a specified theme scenario to address the above-mentioned technical problems, so as to achieve accurate filtering of irrelevant content in ASR output text under a specified theme scenario.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for intelligent content filtering based on a specified theme scenario, the method comprising: S1: Optimize the zero-shot classification model to obtain an optimized zero-shot classification model. The zero-shot classification model performs topic classification based on the constructed knowledge graph prompt template, which includes topic tags and keywords matching the topic tags. S2: The optimized zero-shot classification model is used to classify the first ASR output text into topics, resulting in the first irrelevant content text; S3: Construct a dynamic lexicon, calculate the correlation between the first ASR output text and the dynamic lexicon, filter out the second irrelevant content text based on the correlation, the dynamic lexicon is updated synchronously with the keywords in the knowledge graph prompt template, and is used to fine-tune the optimized zero-shot classification model at preset times; S4: Compare the first irrelevant content text and the second irrelevant content text for duplicate content to obtain the third irrelevant content text; S5: Remove the third irrelevant content text from the first ASR output text to obtain the filtered first ASR output text.

[0008] Preferably, the zero-shot classification model is optimized, including: The second ASR output text is segmented by sentence and the speaker's identity is marked; The constructed knowledge graph prompt template is used to divide each sentence in the second ASR output text into topics, and the preliminary topic labels corresponding to each sentence are obtained. Based on the timestamps of the statements and the continuity of the initial topic tags, the second ASR output text is divided into several segments; Statements with similar meanings within the same segment are classified as positive samples, while statements from different segments or with different meanings are classified as negative samples. Based on the classified positive and negative samples, the zero-shot classification model is fine-tuned using a contrastive loss function to obtain an optimized zero-shot classification model.

[0009] Preferably, the contrastive loss function is calculated as follows:

[0010] in, Indicates comparative loss, This indicates the embedding representation of the current statement. Indicates and Embedded representation of semantically similar positive sample statements The embedding representation of negative sample statements. Represents the cosine similarity function. This represents a temperature hyperparameter used to adjust the sharpness of the distribution. This indicates the statement encoding function. This represents the natural exponential function. This represents all negative sample statements. Indicates the first One negative sample statement.

[0011] Preferably, step S2 includes: An optimized zero-shot classification model is used to classify the first ASR output text into topics, resulting in irrelevant content topic tags, irrelevant content text, and confidence scores. The confidence scores represent the relevance probability between the irrelevant content text and the irrelevant content topic tags. Irrelevant content text with a confidence level higher than a preset threshold is filtered out to obtain the first irrelevant content text.

[0012] Preferably, step S3 includes: A dynamic lexicon is constructed, and the TF-IDF algorithm is used to calculate the correlation between the first ASR output text and the dynamic lexicon; Texts with a relevance lower than a preset threshold in the first ASR output text are filtered out to obtain the second irrelevant content text.

[0013] Preferably, constructing a dynamic lexicon includes: Keyword extraction is performed on the first ASR output text to obtain candidate keywords; Determine whether the candidate keyword is in the dynamic thesaurus. If it is, do not perform a three-dimensional relevance evaluation on the candidate keyword. If it is not, perform a three-dimensional relevance evaluation on the candidate keyword. The candidate keywords are evaluated using a three-dimensional correlation test to obtain an evaluation score. If the evaluation score is greater than the first preset score, the candidate keyword is added to the dynamic lexicon. If the evaluation score is less than the first preset score but greater than the second preset score, the candidate keyword is manually reviewed. If the evaluation score is less than the second preset score, the candidate keyword is not added to the dynamic lexicon.

[0014] Preferably, the candidate keywords obtained include: The N-gram model and TF-IDF algorithm were used to extract high-frequency words from the first ASR output text to obtain high-frequency keywords. The KeyBERT model was used to extract keywords from the first ASR output text to obtain zero-sample keywords; High-frequency keywords and zero-sample keywords are combined to obtain candidate keywords.

[0015] Preferably, the merging includes: Stop words and generic words are removed from high-frequency keywords and zero-sample keywords to obtain the first candidate keywords; Retain keywords from the first candidate keywords whose repetition frequency ranking is lower than the preset ranking among high-frequency keywords or zero-sample keywords, and obtain the second candidate keywords; Remove duplicate words from the second candidate keywords and merge keywords with similar meanings to obtain candidate keywords.

[0016] Preferably, the three-dimensional correlation assessment includes: Calculate the cosine similarity between candidate keywords and a pre-defined core thesaurus for a specified topic scenario; Calculate the co-occurrence probability of candidate keywords with action verbs in a specified topic scenario during the dialogue; The business terms corresponding to the candidate keywords are matched with the preset business dictionary to obtain the preset weight values; The evaluation score is calculated based on cosine similarity, co-occurrence probability, and preset weight values.

[0017] Preferably, if a keyword in the dynamic thesaurus is not used for topic classification within a preset number of days, the keyword will be removed from the dynamic thesaurus.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a method for intelligent content filtering based on a specified topic scenario. It optimizes the zero-shot classification model by using a constructed knowledge graph prompt template as prompt words. The optimized zero-shot classification model effectively improves classification accuracy. A continuously updated dynamic lexicon is constructed, which serves two purposes: firstly, for parallel content filtering in two stages, only content that meets the filtering requirements in both the zero-shot classification stage and the dynamic lexicon matching stage is filtered, minimizing false filtering; secondly, the dynamic lexicon is synchronously updated with keywords in the knowledge graph prompt template, and the zero-shot classification model is trained and optimized periodically, improving the model's adaptability in specific domains. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a method for intelligent content filtering based on a specified theme scenario in one embodiment; Figure 2 This is a schematic diagram illustrating the zero-shot classification model optimization of a method for intelligent content filtering based on a specified topic scenario in one embodiment. Figure 3 This is a schematic diagram of the three-dimensional correlation evaluation process of a method for intelligent content filtering based on a specified topic scenario in one embodiment; Figure 4 This is a schematic diagram of the system architecture of a method for intelligent content filtering based on a specified theme scenario in one embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] Example 1 like Figure 1 As shown, this embodiment proposes a method for intelligent content filtering based on a specified theme scenario, the method including: S1: Optimize the zero-shot classification model to obtain an optimized zero-shot classification model. The zero-shot classification model performs topic classification based on the constructed knowledge graph prompt template, which includes topic tags and keywords matching the topic tags. The specific implementation method for this step is as follows: Taking real estate-themed projects as an example, the goal of zero-shot classification model optimization is to improve the zero-shot model's ability to understand fine-grained semantics in the real estate field and avoid misjudging terms such as "unit type" and "usable floor area ratio" by general models. Figure 2 As shown, the zero-shot classification model is optimized by using knowledge graph prompt templates and fine-tuning them through dialogue scenario comparison.

[0022] Knowledge Graph Hint Template: This template extracts core entities and relationships from a project database within a specified thematic scenario, forming a structured graph. Taking a real estate project database as an example, entities include: apartment layouts, pricing strategies, competitor comparisons, and customer objections. Relationships related to apartment layouts include: number of bedrooms, number of bathrooms, and usable floor area ratio. Relationships related to pricing strategies include: discount rates, payment methods, and competitor benchmarking. Relationships triggered by customer objections include: transportation amenities, school district resources, and delivery standards. After the knowledge graph is constructed, the graph relationships need to be converted into zero-shot classification hints to enhance domain semantic awareness. This is the knowledge graph hint template, which includes multiple topic tags such as apartment layout description, price negotiation, competitor comparison, customer needs analysis, and objection response. In text converted from ASR (Automatic Sentence Recognition), mentioning keywords like "three bedrooms and two living rooms" or "usable floor area ratio" can classify it as an apartment layout description; discussing keywords like "discount" or "loan ratio" can classify it as price negotiation; and mentioning keywords like "XX apartment comparison" can classify it as competitor comparison.

[0023] Dialogue Structure Fine-tuning: 500,000 unlabeled dialogues between real estate consultants and clients were collected as fine-tuning data and input into a zero-shot classification model. Unsupervised contrastive learning was used to train the model to distinguish the semantic weights of consultant (proactive sales) and client (needs / concerns) statements. A Transformer encoder was used to capture cross-statement dependencies (such as clients repeatedly asking about apartment layouts, which were marked as high-weight segments). A contrastive learning loss function was used during training to bring semantically similar statement representations closer together and push away irrelevant statement representations. This loss function compares the similarity of positive and negative samples, enabling the model to learn to distinguish statements of different semantic types and improve the representation quality of the dialogue structure.

[0024] The specific steps are as follows: Input preprocessing: The ASR-processed data is segmented by statement and the speaker's identity (consultant / client) is marked. Knowledge graph topic classification: Using the constructed knowledge graph prompt template, zero-sample topic classification is performed on each statement to obtain preliminary topic tags (such as "apartment type description", "price negotiation", etc.). Dialogue segmentation: Based on statement timestamps and topic continuity, the dialogue is divided into several semantically coherent segments; Positive and negative sample construction: positive samples are pairs of sentences with similar semantics within the same segment, and negative samples are pairs of sentences from different segments or with unrelated semantics; Contrastive learning fine-tuning: The pre-trained language model is fine-tuned using the loss function described above to optimize sentence representation; Output Optimization Topic: The fine-tuned model reclassifies the sentences to obtain more accurate topic tags, providing a basis for subsequent keyword extraction.

[0025] After the zero-shot classification model was optimized, the topic classification accuracy on the test set increased from 74% to 89% of the general model, and the term misclassification rate decreased by 62%.

[0026] S2: The optimized zero-shot classification model is used to classify the first ASR output text into topics, resulting in the first irrelevant content text; The specific implementation method for this step is as follows: An optimized zero-shot classification model is used to classify the topics in the first ASR output text, obtaining the zero-shot classification results and the probability distribution (confidence score) for each category. The optimized zero-shot classification model is a multi-class model, and the output is the relevance probability (confidence score) of each category. Taking real estate as an example, relevant content categories include: price, apartment type, location, orientation, floor, garage, etc., while irrelevant content categories include: food, entertainment, retail, manufacturing, medical, etc. Each category outputs a confidence score and a preset threshold. The first ASR output text is then filtered in one stage, comparing the confidence score of irrelevant content with the threshold. If the probability is higher than the set threshold, it is filtered out, resulting in the first irrelevant content text.

[0027] S3: Construct a dynamic lexicon, calculate the correlation between the first ASR output text and the dynamic lexicon, filter out the second irrelevant content text based on the correlation, the dynamic lexicon is updated synchronously with the keywords in the knowledge graph prompt template, and is used to fine-tune the optimized zero-shot classification model at preset times; The specific implementation method for this step is as follows: Taking real estate projects as an example, in real estate marketing scenarios, the conversations between sales consultants and clients are complex and diverse, containing both explicit real estate terminology and a large amount of peripheral content related to real estate (such as surrounding amenities, school district policies, and transportation planning). To ensure that no potential business opportunities are missed, a two-stage independent filtering mechanism is needed. This involves adding a second stage of filtering; only conversations consistently judged as "irrelevant content" in both stages will be filtered out; otherwise, they will be retained, maximizing the integrity of business information. Figure 4 As shown, an updatable dynamic lexicon was constructed as the second path, working in synergy with the first zero-shot classification path to jointly improve system performance. The core role of the dynamic lexicon in the second path is to provide continuously optimized domain knowledge support for the zero-shot classification engine of the first path. Specifically: Feedback optimization mechanism: New keywords mined from massive dialogues (such as new policy terms "new housing provident fund policy", new project name "Xinghan TOD") will be added to the dynamic thesaurus in real time. In order to ensure that no potential business opportunities are lost, a two-stage independent filtering mechanism is required. The dynamic thesaurus will be used in real time for the second filtering after zero-sample classification filtering.

[0028] Enhanced classification capabilities: The updated dynamic lexicon will be updated synchronously with the keywords in the knowledge graph prompt template, enabling the zero-shot classification model to recognize these emerging terms. Since the prompt words that rely on the knowledge graph during the reasoning process will affect the reasoning efficiency as the capacity of the knowledge graph increases, the updated dynamic lexicon will also be used as a training set to periodically fine-tune the BERT zero-shot classification model, build a new zero-shot classification model for classification, form a business closed loop, and improve the accuracy and performance of the model.

[0029] Closed-loop optimization system: Two paths form a continuously optimizing closed loop. The second path updates the lexicon, and the enhanced lexicon improves the filtering accuracy of the first path. This design ensures that the system can adapt to the language changes in real estate marketing scenarios, breaking through the limitations of traditional static lexicons.

[0030] By constructing a dynamic lexicon, the first ASR output text is filtered in two stages. The first ASR output text and the dynamic lexicon are compared using the TF-IDF algorithm to calculate the relevance. Specifically, the first ASR output text is first preprocessed and segmented. The preprocessed and segmented text is compared with the dynamic lexicon to count the matching words related to real estate. The counted words are divided by the total number of words in the irrelevant content text to obtain the relevant word density. When the relevant word density is lower than a preset threshold, the irrelevant content text is determined to need to be filtered to obtain the second irrelevant content text.

[0031] S4: Compare the first irrelevant content text and the second irrelevant content text for duplicate content to obtain the third irrelevant content text; The specific implementation method for this step is as follows: Only conversations that are consistently deemed irrelevant by both stages will be filtered; otherwise, they will be retained to ensure the integrity of business information to the greatest extent possible.

[0032] S5: Remove the third irrelevant content text from the first ASR output text to obtain the filtered first ASR output text.

[0033] Example 2 This embodiment further explains the method for intelligent content filtering based on a specified theme scenario proposed in Embodiment 1.

[0034] Optimization of the zero-shot classification model includes: The second ASR output text is segmented by sentence and the speaker's identity is marked; The constructed knowledge graph prompt template is used to divide each sentence in the second ASR output text into topics, and the preliminary topic labels corresponding to each sentence are obtained. Based on the timestamps of the statements and the continuity of the initial topic tags, the second ASR output text is divided into several segments; Statements with similar meanings within the same segment are classified as positive samples, while statements from different segments or with different meanings are classified as negative samples. Based on the classified positive and negative samples, the zero-shot classification model is fine-tuned using a contrastive loss function to obtain an optimized zero-shot classification model.

[0035] The expression for calculating the contrastive loss function is as follows:

[0036] in, Indicates comparative loss, This indicates the embedding representation of the current statement. Indicates and Embedded representation of semantically similar positive sample statements The embedding representation of negative sample statements. Represents the cosine similarity function. This represents a temperature hyperparameter used to adjust the sharpness of the distribution. This indicates the statement encoding function. This represents the natural exponential function. This represents all negative sample statements. Indicates the first One negative sample statement.

[0037] Example 3 This embodiment further explains the method for intelligent content filtering based on a specified theme scenario proposed in Embodiment 1.

[0038] Step S2 includes: An optimized zero-shot classification model is used to classify the first ASR output text into topics, resulting in irrelevant content topic tags, irrelevant content text, and confidence scores. The confidence scores represent the relevance probability between the irrelevant content text and the irrelevant content topic tags. Irrelevant content text with a confidence level higher than a preset threshold is filtered out to obtain the first irrelevant content text.

[0039] Step S3 includes: A dynamic lexicon is constructed, and the TF-IDF algorithm is used to calculate the correlation between the first ASR output text and the dynamic lexicon; Texts with a relevance lower than a preset threshold in the first ASR output text are filtered out to obtain the second irrelevant content text.

[0040] Building a dynamic thesaurus includes: Keyword extraction is performed on the first ASR output text to obtain candidate keywords; Determine whether the candidate keyword is in the dynamic thesaurus. If it is, do not perform a three-dimensional relevance evaluation on the candidate keyword. If it is not, perform a three-dimensional relevance evaluation on the candidate keyword. The candidate keywords are evaluated using a three-dimensional correlation test to obtain an evaluation score. If the evaluation score is greater than the first preset score, the candidate keyword is added to the dynamic lexicon. If the evaluation score is less than the first preset score but greater than the second preset score, the candidate keyword is manually reviewed. If the evaluation score is less than the second preset score, the candidate keyword is not added to the dynamic lexicon.

[0041] The candidate keywords obtained include: The N-gram model and TF-IDF algorithm were used to extract high-frequency words from the first ASR output text to obtain high-frequency keywords. The KeyBERT model was used to extract keywords from the first ASR output text to obtain zero-sample keywords; High-frequency keywords and zero-sample keywords are combined to obtain candidate keywords.

[0042] The specific implementation method for this step is as follows: Taking a real estate-themed project as an example, in order to overcome the limitations of a static thesaurus and dynamically capture emerging terms in marketing scenarios (such as "relaxation of purchase restrictions" and "trade-in policy"), a dual-path collaborative extraction of new keywords is adopted. Specifically, the N-gram model and TF-IDF algorithm are used to extract high-frequency words from the first ASR output text to obtain high-frequency keywords; the KeyBERT model is used to extract keywords from the first ASR output text to obtain zero-sample keywords. The high-frequency keywords and zero-sample keywords are then merged to obtain candidate keywords. The implementation steps are as follows: Input the first ASR output text (divided by sentence) after topic classification and the topic labels output by the zero-shot classification model (such as "price negotiation", "school district demand", etc.); Topic tags are used as KeyBERT Seed Keywords to restrict the semantic direction of keyword extraction. The KeyBERT model is called to extract keywords, and keywords with the same semantics in the consultant's dialogue with the client are merged (e.g., when the client asks "usable floor area ratio," the consultant explains "internal area ratio," merging them into "usable floor area ratio = internal area ratio") to obtain zero-sample keywords. In addition to using the zero-sample-guided KeyBERT path to extract zero-sample keywords, traditional N-gram models and TF-IDF algorithms are also used to extract high-frequency keywords, such as "three-bedroom apartment" and "down payment ratio." Zero-sample keywords and high-frequency keywords are merged, and stop words (such as "this" and "then") and generic words (such as "okay" and "understand") are removed. Candidate words that rank in the top 10 in at least one path are retained. Duplicate words are deduplicated, and semantically similar words are merged. Finally, candidate keywords are obtained. For example, the traditional path outputs: "three-bedroom apartment", "down payment ratio", "school district"; the zero-sample path outputs: "relaxed purchase restrictions", "trade-in", "usable floor area ratio"; after merging, it becomes: "three-bedroom apartment", "down payment ratio", "school district", "relaxed purchase restrictions", "trade-in", "usable floor area ratio". After obtaining candidate keywords, it is determined whether the candidate keyword is in the dynamic thesaurus. If it is not in the dynamic thesaurus, a three-dimensional relevance evaluation is performed, and then a decision is made on whether to add it to the dynamic thesaurus.

[0043] The merger includes: Stop words and generic words are removed from high-frequency keywords and zero-sample keywords to obtain the first candidate keywords; Retain keywords from the first candidate keywords whose repetition frequency ranking is lower than the preset ranking among high-frequency keywords or zero-sample keywords, and obtain the second candidate keywords; Remove duplicate words from the second candidate keywords and merge keywords with similar meanings to obtain candidate keywords.

[0044] The three-dimensional correlation assessment includes: Calculate the cosine similarity between candidate keywords and a pre-defined core thesaurus for a specified topic scenario; Calculate the co-occurrence probability of candidate keywords with action verbs in a specified topic scenario during the dialogue; The business terms corresponding to the candidate keywords are matched with the preset business dictionary to obtain the preset weight values; The evaluation score is calculated based on cosine similarity, co-occurrence probability, and preset weight values.

[0045] The specific implementation method for this step is as follows: Taking real estate-themed projects as an example, in order to quantify the true relevance of new keywords to real estate marketing and avoid misclassifying irrelevant words (such as "hot weather" which may be high-frequency but not crucial), a three-dimensional relevance evaluation model is needed to evaluate candidate keywords. The evaluation process is as follows: Figure 3 As shown, the specific three dimensions are as follows: Semantic similarity (40%): Calculate the cosine similarity between candidate words and the core real estate vocabulary (such as "building", "unit type", "price"); use a pre-trained Chinese embedding model (such as text2vec-base-chinese); threshold setting: greater than or equal to 0.65 is considered a valid association.

[0046] Context Dependency (30%): Analyze the co-occurrence probability of candidate words with marketing action verbs ("buy", "compare", "recommend", etc.) in the dialogue; based on sliding window co-occurrence statistics (window size = 5 words); threshold setting: greater than or equal to 0.6 is considered a valid association.

[0047] Business weight (30%): Matches the enterprise's preset business dictionary (project name, regional policies, company's special services, etc.); manually preset weight values ​​(e.g., "Xinghuicheng" = 0.95, "New Housing Provident Fund Policy" = 0.85); threshold setting: greater than or equal to 0.7 is a valid association.

[0048] As shown in Table 1, when the three-dimensional comprehensive score is greater than or equal to 0.7, the candidate keywords are added to the dynamic thesaurus; when the three-dimensional comprehensive score is greater than or equal to 0.5 and less than 0.7, the candidate keywords are manually reviewed before it is determined whether to add them to the dynamic thesaurus; when the three-dimensional comprehensive score is less than 0.5, the candidate keywords are discarded.

[0049]

[0050] Table 1 If a keyword in the dynamic thesaurus is not used for topic classification within a preset number of days, the keyword will be removed from the dynamic thesaurus.

Claims

1. A method for intelligent content filtering based on a specified theme scenario, characterized in that, include: S1: Optimize the zero-shot classification model to obtain an optimized zero-shot classification model. The zero-shot classification model performs topic classification based on the constructed knowledge graph prompt template, which includes topic tags and keywords matching the topic tags. S2: The optimized zero-shot classification model is used to classify the first ASR output text into topics, resulting in the first irrelevant content text; S3: Construct a dynamic lexicon, calculate the correlation between the first ASR output text and the dynamic lexicon, filter out the second irrelevant content text based on the correlation, the dynamic lexicon is updated synchronously with the keywords in the knowledge graph prompt template, and is used to fine-tune the optimized zero-shot classification model at preset times; S4: Compare the first irrelevant content text and the second irrelevant content text for duplicate content to obtain the third irrelevant content text; S5: Remove the third irrelevant content text from the first ASR output text to obtain the filtered first ASR output text.

2. The method for intelligent content filtering based on a specified theme scenario according to claim 1, characterized in that, Optimization of the zero-shot classification model includes: The second ASR output text is segmented by sentence and the speaker's identity is marked; The constructed knowledge graph prompt template is used to divide each sentence in the second ASR output text into topics, and the preliminary topic labels corresponding to each sentence are obtained. Based on the timestamps of the statements and the continuity of the initial topic tags, the second ASR output text is divided into several segments; Statements with similar meanings within the same segment are classified as positive samples, while statements from different segments or with different meanings are classified as negative samples. Based on the classified positive and negative samples, the zero-shot classification model is fine-tuned using a contrastive loss function to obtain an optimized zero-shot classification model.

3. The method for intelligent content filtering based on a specified theme scenario according to claim 2, characterized in that, The expression for calculating the contrastive loss function is as follows: in, Indicates comparative loss, This indicates the embedding representation of the current statement. Indicates and Embedded representation of semantically similar positive sample statements The embedding representation of negative sample statements. Represents the cosine similarity function. This represents a temperature hyperparameter used to adjust the sharpness of the distribution. This indicates the statement encoding function. This represents the natural exponential function. This represents all negative sample statements. Indicates the first One negative sample statement.

4. The method for intelligent content filtering based on a specified theme scenario according to claim 1, characterized in that, Step S2 includes: An optimized zero-shot classification model is used to classify the first ASR output text into topics, resulting in irrelevant content topic tags, irrelevant content text, and confidence scores. The confidence scores represent the relevance probability between the irrelevant content text and the irrelevant content topic tags. Irrelevant content text with a confidence level higher than a preset threshold is filtered out to obtain the first irrelevant content text.

5. The method for intelligent content filtering based on a specified theme scenario according to claim 1, characterized in that, Step S3 includes: A dynamic lexicon is constructed, and the TF-IDF algorithm is used to calculate the correlation between the first ASR output text and the dynamic lexicon; Texts with a relevance lower than a preset threshold in the first ASR output text are filtered out to obtain the second irrelevant content text.

6. The method for intelligent content filtering based on a specified theme scenario according to claim 1, characterized in that, Building a dynamic thesaurus includes: Keyword extraction is performed on the first ASR output text to obtain candidate keywords; Determine whether the candidate keyword is in the dynamic thesaurus. If it is, do not perform a three-dimensional relevance evaluation on the candidate keyword. If it is not, perform a three-dimensional relevance evaluation on the candidate keyword. The candidate keywords are evaluated using a three-dimensional correlation test to obtain an evaluation score. If the evaluation score is greater than the first preset score, the candidate keyword is added to the dynamic lexicon. If the evaluation score is less than the first preset score but greater than the second preset score, the candidate keyword is manually reviewed. If the evaluation score is less than the second preset score, the candidate keyword is not added to the dynamic lexicon.

7. The method for intelligent content filtering based on a specified theme scenario according to claim 6, characterized in that, The candidate keywords obtained include: The N-gram model and TF-IDF algorithm were used to extract high-frequency words from the first ASR output text to obtain high-frequency keywords. The KeyBERT model was used to extract keywords from the first ASR output text to obtain zero-sample keywords; High-frequency keywords and zero-sample keywords are combined to obtain candidate keywords.

8. The method for intelligent content filtering based on a specified theme scenario according to claim 7, characterized in that, The merger includes: Stop words and generic words are removed from high-frequency keywords and zero-sample keywords to obtain the first candidate keywords; Retain keywords from the first candidate keywords whose repetition frequency ranking is lower than the preset ranking among high-frequency keywords or zero-sample keywords, and obtain the second candidate keywords; Remove duplicate words from the second candidate keywords and merge keywords with similar meanings to obtain candidate keywords.

9. The method for intelligent content filtering based on a specified theme scenario according to claim 6, characterized in that, The three-dimensional correlation assessment includes: Calculate the cosine similarity between candidate keywords and a pre-defined core thesaurus for a specified topic scenario; Calculate the co-occurrence probability of candidate keywords with action verbs in a specified topic scenario during the dialogue; The business terms corresponding to the candidate keywords are matched with the preset business dictionary to obtain the preset weight values; The evaluation score is calculated based on cosine similarity, co-occurrence probability, and preset weight values.

10. The method for intelligent content filtering based on a specified theme scenario according to claim 1, characterized in that, If a keyword in the dynamic thesaurus is not used for topic classification within a preset number of days, the keyword will be removed from the dynamic thesaurus.

Images