Service metadata-oriented generative semantic fusion recommendation method

Abstract

The invention provides a service metadata-oriented generative semantic fusion recommendation method, and belongs to the field of recommendation systems. The method comprises the following steps: S1, processing an original data set containing metadata pairs and relevancy marks to generate a high-order metadata training set containing semantic enhancement data and difficult negative samples; s2, performing multi-task training and iterative optimization on a pre-trained Embedding model based on the high-order metadata training set to obtain a field adaptive semantic understanding model; and S3, performing coarse-grained retrieval on the query metadata by utilizing the field adaptive semantic understanding model, performing fine-grained screening and reordering on a retrieval result by utilizing a large language model in combination with a thinking chain strategy, and outputting a final recommendation result. According to the method, through a multi-dimensional deep semantic field analysis technology, in combination with deep understanding of a large language model on professional terms and complex contexts, the context expression ability of metadata in the vertical field is improved, and the distinguishing ability of similar but unrelated metadata is optimized.

Description

Technical Field

[0001] This invention relates to the field of recommender systems, and more specifically to a generative semantic fusion recommender method oriented towards service metadata. Background Technology

[0002] Recommendation engines, as a core technology for information filtering and content distribution, have been widely applied in the data service field. Their core task is to analyze user query input and accurately identify and recommend highly relevant content from massive, heterogeneous service metadata. In recent years, with the advancement of natural language processing technology, recommendation methods based on text semantic understanding have become a research hotspot. Compared with traditional methods based on collaborative filtering or rule matching, semantic recommendation can deeply analyze the connotation of text, theoretically better handle the new item (cold start) problem, and improve the interpretability and accuracy of recommendations.

[0003] However, the core challenge in applying semantic technology to service metadata recommendation lies in the accurate understanding and quantification of the deep semantic relationships between metadata. Service metadata typically includes structured or semi-structured text such as interface descriptions, functional specifications, and technical parameters. It is highly specialized, terminologically dense, and context-dependent. Traditional text similarity calculation methods, such as TF-IDF based on word frequency statistics or BM25 based on probabilistic models, while computationally efficient and simple to implement, are essentially shallow matching based on the bag-of-words model. These methods severely lack the ability to understand the "semantic field" in which words exist, and cannot effectively handle polysemy, synonyms, and the specific meanings of technical terms in particular contexts. This results in significantly limited recommendation accuracy when facing complex business contexts.

[0004] To overcome the limitations of shallow semantic representations, deep learning-based text vectorization models (such as Word2Vec, GloVe, and more advanced pre-trained language models like BERT and RoBERTa) have been widely introduced. These models, through pre-training on large-scale corpora, can generate context-sensitive word vectors (embeddings) containing rich semantic information, thereby significantly improving the model's ability to capture the semantics of general languages. Nevertheless, directly applying these general models to service metadata recommendation scenarios in vertical domains still faces three prominent problems: First, general pre-trained corpora are insufficient to cover the massive amounts of specialized terms and abbreviations with specific meanings in numerous vertical domains (such as SOA service architecture, vehicle networking protocols, bioinformatics, and financial risk control), leading to deviations in the model's interpretation of the true intent of metadata; second, implicit knowledge such as industry background knowledge, business logic constraints, and compliance requirements is usually not explicitly encoded into the text and is difficult to learn automatically through general models, while this knowledge is crucial for judging the business relevance between metadata; finally, in real-time recommendation scenarios facing massive service libraries, a balance needs to be struck between extremely high recall speed and fine semantic ranking accuracy, but complex deep models often bring high computational overhead, making it difficult to directly meet the needs of high-concurrency, low-latency online services.

[0005] In practical applications, existing technologies have significant shortcomings in several key areas. At the semantic understanding level, traditional statistical methods and general neural network models are insufficient in modeling specialized terminology and complex business logic, resulting in limited discriminative power in the generated semantic representations. At the model training level, high-quality negative samples are crucial for advanced training paradigms such as contrastive learning, but existing methods often employ random sampling or strategies based on fixed similarity thresholds to construct negative samples. This strategy struggles to generate "hard negative samples" that are "similar but irrelevant" to positive samples, preventing the model from learning subtle but crucial semantic differences and thus affecting its final discriminative ability. At the retrieval and ranking level, mainstream methods typically transform the recommendation problem into a vector retrieval problem: first, obtain the query embedding; then, recall the candidate set through approximate nearest neighbor search; and finally, rank the samples based on simple metrics such as cosine similarity. This approach has two major drawbacks: First, the embedding representation itself may contain noise or fail to fully capture business relevance, and relying solely on vector distance sorting can easily introduce semantically relevant but business-irrelevant "interference items"; Second, when the service metadata database is extremely large, accurate full-database retrieval is computationally expensive, while using approximate retrieval algorithms may result in a loss of accuracy, creating a contradiction between effectiveness and efficiency.

[0006] In summary, current semantic recommendation technologies based on service metadata have significant shortcomings in areas such as deep semantic field parsing, high-quality construction of difficult samples, and retrieval and ranking mechanisms that balance accuracy and efficiency. This directly limits the application effectiveness and practical value of recommendation systems in specialized, high-requirement business scenarios (such as enterprise-level service governance, precise API recommendations, and industry knowledge services). Therefore, there is an urgent need for a new recommendation method that can deeply integrate domain knowledge, finely model semantic relationships, and achieve efficient and accurate retrieval. Summary of the Invention

[0007] The purpose of this invention is to address the problems existing in the prior art and provide a generative semantic fusion recommendation method for service metadata. It employs multi-dimensional deep semantic field parsing enhancement technology to parse the syntactic structure of input metadata and enhance its semantics in conjunction with context; it optimizes the selection of negative metadata through a dynamic hard negative metadata mining strategy, ensuring that the semantic difference between negative and positive metadata remains within a reasonable range; and it utilizes deep reasoning technology guided by contrastive knowledge to perform semantic field filtering and ranking optimization on the candidate set after the initial candidate set is recalled by the embedding model, leveraging the contextual understanding capabilities of a large model.

[0008] This invention is achieved through the following technical solution:

[0009] This invention provides a generative semantic fusion recommendation method based on service metadata, comprising:

[0010] S1: Process the original dataset containing metadata pairs and relevance labels to generate a high-order metadata training set containing semantically enhanced data and hard negative samples;

[0011] S2: Based on the high-order metadata training set, perform multi-task training and iterative optimization on the pre-trained Embedding model to obtain a domain-adaptive semantic understanding model;

[0012] S3: Utilize a domain-adaptive semantic understanding model to perform coarse-grained retrieval of query metadata, and use a large language model combined with a thought chain strategy to perform fine-grained filtering and reordering of the retrieval results, outputting the final recommendation results.

[0013] Furthermore, S1 includes:

[0014] S101: Input the original dataset, divide the original dataset into a training set and a validation set, and construct a recall set consisting of all metadata;

[0015] S102: Use the Large Language Model (LLM) to perform domain knowledge parsing and semantic augmentation on the metadata in the original dataset to generate a semantically enhanced dataset;

[0016] S103: Based on the dynamic threshold partitioning strategy, the sample labels are converted into positive and negative samples, and based on the pre-trained Embedding model and the set similarity threshold range, difficult negative samples are mined from the recall set and combined to generate a high-order metadata training set.

[0017] Furthermore, the dynamic threshold division strategy in S103 includes:

[0018] Based on the original relevance label set, the fixed interval between adjacent labeled data in the set is k, and the set S is represented as follows:

[0019] S={s i |1≤i≤n,n≥2,s i -s i-1 =k,k∈C}

[0020] Where i represents the labeled data, k represents the interval between adjacent labeled data, and n represents the number of labeled data;

[0021] Generate candidate thresholds, which are dynamic thresholds, as follows:

[0022] t′=t0+k*(i-1)

[0023] Where t′ is the candidate threshold, and t0 is the initial threshold, t0 = min(s1,...,s n ), 1≤i≤n;

[0024] The training set is divided and the model is trained using different candidate thresholds. The model performance is evaluated on the validation set, and the candidate threshold with the best comprehensive score is selected as the final splitting threshold t.

[0025] The sample pairs are labeled as positive or negative samples based on the threshold t.

[0026] Furthermore, the operation of mining difficult negative samples in S103 is as follows: for each metadata x in the training set i Calculate the cosine approximation of the metadata in the recall set with all metadata in the training set; i Sort by similarity in descending order; select metadata whose similarity values ​​fall within a preset threshold range to form a dataset x. i The difficult negative sample set.

[0027] Furthermore, S2 includes:

[0028] S201: Based on the semantic augmentation dataset, the pre-trained Embedding model is fine-tuned with all parameters using the domain category labels of metadata to perform multi-class classification tasks;

[0029] S202: Based on the high-order metadata training set, the contrastive learning loss function is used to further train the fine-tuned Embedding model;

[0030] S203: On the validation set, through adaptive iterative optimization, the optimal parameter combination is selected, and the difficult negative sample mining and embedding model training process is iteratively executed until the termination condition is met, thus obtaining a domain adaptive semantic understanding model.

[0031] Furthermore, the full parameter fine-tuning operation in step 201 includes: adding a classification head based on the expanded dataset D′ and fine-tuning using the following multi-class cross-entropy loss function:

[0032]

[0033] Where C represents the total number of categories, y i p represents the one-hot encoding of the actual label. i This represents the Softmax probability output by the model.

[0034] Furthermore, the contrastive learning loss function in S202 is the InfoNCE loss function, with the following expression:

[0035]

[0036] Among them, h i It is the embedding representation of the initial metadata. It is the embedding representation of positive metadata. It is the embedding representation of negative metadata, sim(h) i ,h j ) is the approximation calculation between two embeddings, and τ is the temperature hyperparameter.

[0037] Furthermore, the adaptive iterative optimization of S203 includes: testing different threshold combinations on the validation set through grid search to evaluate macro and micro indicators; setting an early stopping mechanism to terminate the iteration; and finally retraining the model with the full dataset.

[0038] Furthermore, S3 includes:

[0039] S301: Employs a domain-adaptive semantic understanding model to encode query metadata into vector representations and efficiently retrieves a coarse-ranked candidate metadata set, ten times the size of the predetermined recommended metadata, from the recall set using an approximate nearest neighbor search algorithm.

[0040] S302: Construct prompt words containing the coarse-ranked candidate metadata information, domain knowledge, and specific task instructions, and input them into the Large Language Model (LLM);

[0041] S303: Using the large language model LLM, perform deep semantic analysis and relevance judgment on the coarsely ranked candidate metadata, and output the final recommended metadata after filtering and reordering.

[0042] Furthermore, in step S303, the final recommendation metadata of the Large Language Model (LLM) is in a structured format. The final recommendation metadata is parsed using regular expressions to obtain the final recommendation result.

[0043] The beneficial effects of this invention are:

[0044] By employing multi-dimensional deep semantic field parsing technology, combined with a large language model's deep understanding of specialized terminology and complex contexts, the model's ability to express contextual information about metadata in vertical domains is significantly improved. Simultaneously, a high-order metadata construction strategy optimizes data quality and enhances the model's ability to distinguish between similar but irrelevant metadata. During the retrieval phase, a semantic enhancement method incorporating thought chains is used, leveraging the large model's reasoning capabilities to perform secondary filtering and ranking of the retrieved results. This effectively avoids the false recall problem caused by semantic field shifts in traditional methods, ensuring that the recommended results are more aligned with domain requirements.

[0045] Experiments show that this technology significantly outperforms traditional methods in key metrics such as Recall@K, Precision@K, and NDCG@K, maintaining high accuracy even in scenarios with small values ​​like K=3 and K=5, and greatly improving the coverage and ranking priority of relevant metadata. Through deep semantic understanding and optimized ranking strategies, the model achieves high recall, high precision, and stable ranking capabilities in professional scenarios, ensuring the professional reliability of recommendation results while improving user experience and system usability, providing an efficient solution for metadata recommendation in vertical domains. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings:

[0048] The implementation process of this invention can be divided into three core stages: high-order metadata expansion, adaptive model training, and multi-level deep retrieval and recommendation.

[0049] Phase 1: High-level metadata expansion

[0050] Step 1-1, Input and Data Division

[0051] Input the original dataset Where x i and y iThese are service metadata pairs, and the relevance between them is indicated by s i The representation can be a binary value (0 or 1) to indicate whether the similarity is unrelated or related, or a range of real numbers / integers to provide a more nuanced distinction of similarity.

[0052] Dataset The dataset is randomly divided into training and validation sets in a ratio of 7:3 or other proportions (using 7:3 as an example).

[0053] Building a recall set It is by The text collection consisting of service metadata is represented as:

[0054]

[0055] The recall set R is composed of all metadata text x from the original dataset D. i and y i The set after deduplication is structured as follows:

[0056] Structure: Dataset D contains n sample pairs (x... i ,y i ,S i Each sample pair contains two metadata texts x. i and y i The recall set R is the text set formed by merging and deduplicating all these metadata texts.

[0057] Quantitative relation: If all metadata texts are unique, then |R| = 2n; if duplicate texts exist, then |R| < 2n. In practical applications, since different sample pairs may share the same metadata, the actual size of the recall set R is usually less than 2n.

[0058] Preprocessing: The recall set R needs to undergo data cleaning operations (such as deduplication, case neutralization, removal of special characters, etc.) to ensure data quality and computational efficiency.

[0059] Furthermore, recall Data cleaning is required, such as deduplication and conversion to lowercase.

[0060] Steps 1-2: LLM generates data

[0061] LLM semantic augmentation and word order structure decomposition. Multi-dimensional deep semantic field parsing enhancement techniques are used for processing. Metadata in the service metadata is used to enhance the text's representation capabilities and domain adaptability. Leveraging the deep semantic understanding capabilities of LLM, domain knowledge is extracted from the service metadata to construct a semantically enhanced dataset.

[0062]

[0063] where p i represents the domain knowledge generated by the LLM.

[0064] This step endows the original data with richer context and domain connotations.

[0065] Steps 1 - 3, positive and negative sample generation

[0066] (1) Positive and negative metadata segmentation: Based on the dynamic parameter adaptive tuning technology, set the positive and negative metadata thresholds, divide the labels of metadata pairs into relevant (1) and irrelevant (0), construct a dataset in Pairwise format. If the relevance label in is a binary value, then this step is ignored. Since the actual relevance values are usually linear, that is, the numerical gap k between adjacent labeled data is the same, the label set can be expressed as S = {s i ∣1 ≤ i ≤ n, n ≥ 2, s i - s i-1 = k, k ∈ C}.

[0067] Set the initial threshold as t0 = min(s1,..., s n ), and set the optimization span step as k, then the dynamic threshold can be obtained:

[0068] t′ = t0 + k * (i - 1)

[0069] where 1 ≤ i ≤ n.

[0070] For each candidate threshold t’ i , use this partitioning scheme to train the Embedding model, evaluate the model performance on the validation set Dval, calculate key metrics (such as Recall@K, Precision@K, NDCG@K), select the threshold with the best performance metrics on the validation set as the final threshold t, and divide the training set into positive and negative samples according to this threshold. The evaluation criterion can adopt a comprehensive score. For example: Score = α·Recall@5 + β·Precision@5 + γ·NDCG@5, where α, β, γ are weight coefficients and can be adjusted according to business requirements.

[0071] The relevance label for each sample can be converted to:

[0072]

[0073] That is, if s i ≥ t, then mark the sample pair (x i , y i ) as a positive sample (labeled 1); if s i < t, then mark the sample pair (xi ,y i ) are marked as negative samples (label is 0).

[0074] (2) Hard Negative Sample Mining: To improve the model's ability to recognize fine-grained semantic differences, high-order metadata construction techniques are used to mine negative samples. This is achieved by calculating negative samples based on zero-shot embedding models (such as BERT, RoBERTa, etc.) or embedding models of the original text trained using the techniques of this invention. Each metadata text x i With Recall Collection Chinese metadata text r j Cosine approximation:

[0075]

[0076] Wherein, E(x) i ) and E(r j ) represent metadata x respectively i and recall centralized metadata r j Embedding. For Metadata x in the i-th sample i Arranging the approximations in descending order yields... The set of difficult negative data was selected from the data based on the relevance score thresholds p1 and p2. The initial values ​​of p1 and p2 can be 0.90 and 0.95 respectively, and then dynamically adjusted according to the actual task.

[0077] The higher-order meta-dataset obtained based on the i-th sample can be represented as:

[0078]

[0079] Finally, all metadata is integrated to obtain a high-order metadata training set:

[0080]

[0081] Here, n refers to... The quantity.

[0082] In vertical industries (such as SOA services, connected vehicles, etc.), a large number of technical terms lead to many metadata that appear similar on the surface but have different actual meanings. Through the above-mentioned difficult negative sample mining operation, the model can be effectively trained to identify these subtle differences.

[0083] The second stage is adaptive model training.

[0084] Step 2-1, Fine-tuning of all parameters

[0085] During the model training phase, the model is first trained based on the semantically augmented dataset. The pre-trained embedding model is fine-tuned with all parameters. A classification head (default fully connected layer) is added to the model, and all parameters (including the pre-trained layer) are optimized to give it stronger generalization ability in understanding the semantics of service metadata in a specific domain.

[0086] In this invention, service metadata is categorized by domain type. These categories originate from the different domain types to which the metadata in dataset D belongs, and the metadata in the dataset is labeled with category tags. In vertical domain recommendation scenarios, service metadata typically belongs to different business or functional categories. For example, SOA service domains include: data processing, API call, and message transmission.

[0087] For categorical data, the multi-class cross-entropy loss function is used:

[0088]

[0089] Where C represents the total number of categories, y i p represents the one-hot encoding of the actual label. i This represents the Softmax probability output by the model.

[0090] Classification tasks force models to learn more discriminative feature representations, enabling better separation of metadata from different categories in the vector space. Furthermore, models fine-tuned through classification tasks already possess strong domain semantic understanding capabilities, allowing for better subsequent contrastive learning training. Compared to directly performing contrastive learning, warming up with a multi-classification task first prevents the model from converging to local optima in the early stages of contrastive learning.

[0091] Step 2-2, Comparative Learning and Training

[0092] Fine-tuning of the model parameters after cross-entropy training is performed, and fine-grained semantic differences are captured through comparative learning.

[0093] Specifically, a training set containing higher-order metadata is obtained based on the approximation label threshold t and the relevance score thresholds p1 and p2. The Embedding model is trained using contrastive training and semantically tuned based on a contrastive loss function (such as InfoNCE Loss) to further optimize the robustness and accuracy of the model in contextual understanding.

[0094] The Loss function is as follows:

[0095]

[0096] Among them, h i It is the embedding representation of the initial metadata. It is the embedding representation of positive metadata. It is the embedding representation of negative metadata, sim(h) i ,h j ) is an approximation operation between two embeddings, usually using the Cosine approximation, where τ is the temperature hyperparameter.

[0097] Steps 2-3: Adaptive iterative optimization of parameters

[0098] Grid Search in the validation set Test different threshold combinations The learned embedding model is used to measure Recall@K, Precision@K, and NDCG@K using both macro and micro methods to evaluate its predictive performance and obtain the optimal parameter combination. Then, the entire dataset is used... After augmentation with higher-order samples, retraining is performed. The calculation of Recall@K, etc., needs to consider the ranking factor (the higher the ranking, the larger the value). Here, S is used. i The size of the label in (D) is used as the sorting criterion.

[0099] The trained model can then be used as an embedding model for approximation calculation in higher-order metadata augmentation, further enhancing its ability to distinguish difficult negative metadata and improving its quality. The embedding model is then trained again based on new data. This step is an adaptive iterative optimization loop with explicit triggering and termination conditions, and it provides an early stopping mechanism: if the validation set performance improvement is less than a threshold ε (e.g., ε = 0.5%) for M consecutive iterations (e.g., M = 5 or 10) or the preset maximum number of iterations Tmax (usually set to 3-5 iterations) is reached, iteration stops. This adaptive training method can significantly improve the training effect of contrastive learning.

[0100] The third stage involves multi-level, in-depth search and recommendation.

[0101] Step 3-1, coarse-grained retrieval

[0102] In the retrieval phase, a deep reasoning technique guided by comparative knowledge is adopted. First, a trained Embedding model is used in combination with the nearest neighbor algorithm (KNN algorithm) to efficiently retrieve the metadata of the top 10K most similar candidate services, forming a coarse candidate set.

[0103] Step 3-2, Fine-grained screening and reordering

[0104] Subsequently, the metadata information of the coarsely ranked candidate set is input into the large language model LLM. Combined with the large model-driven Chain-of-Thought (CoT) strategy, the candidate set is deeply filtered and re-ranked to obtain the top K results. Finally, the top K most relevant and accurate recommendation results are output.

[0105] Specifically, LM filtering and rearrangement employs semantic reasoning technology based on Chain-of-Thought (CoT), and the specific implementation method is as follows:

[0106] 1. Create a prompt. The prompt should include candidate sets (to be filled), domain knowledge (to be filled), etc., to clearly define the task as filtering and rearranging, and specify the output format (such as JSON). LLM will then output text in a fixed format.

[0107] 2. Input candidate set information and domain knowledge;

[0108] 3. Parse the output using regular expressions to obtain the Top-K.

[0109] The above technical solution is only one embodiment of the present invention. For those skilled in the art, based on the principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the technical solutions described in the specific embodiments of the present invention. Therefore, the foregoing description is only a preferred option and is not restrictive.

Claims

1. A generative semantic fusion recommendation method oriented towards service metadata, characterized in that, The method includes: S1: Process the original dataset containing metadata pairs and relevance labels to generate a high-order metadata training set containing semantically enhanced data and hard negative samples; S2: Based on the high-order metadata training set, perform multi-task training and iterative optimization on the pre-trained Embedding model to obtain a domain-adaptive semantic understanding model; S3: Utilize a domain-adaptive semantic understanding model to perform coarse-grained retrieval of query metadata, and use a large language model combined with a thought chain strategy to perform fine-grained filtering and reordering of the retrieval results, outputting the final recommendation results.

2. The generative semantic fusion recommendation method for service metadata as described in claim 1, characterized in that, S1 includes: S101: Input the original dataset, divide the original dataset into a training set and a validation set, and construct a recall set consisting of all metadata; S102: Use the Large Language Model (LLM) to perform domain knowledge parsing and semantic augmentation on the metadata in the original dataset to generate a semantically enhanced dataset; S103: Based on the dynamic threshold partitioning strategy, the sample labels are converted into positive and negative samples, and based on the pre-trained Embedding model and the set similarity threshold range, difficult negative samples are mined from the recall set and combined to generate a high-order metadata training set.

3. The generative semantic fusion recommendation method for service metadata as described in claim 2, characterized in that, The dynamic threshold division strategy in S103 includes: Based on the original relevance label set, the fixed interval between adjacent labeled data in the set is k, and the set S is represented as follows: S={s i ∣1≤i≤n,n≥2,s i -s i-1 =k,k∈C} Where i represents the labeled data, k represents the interval between adjacent labeled data, and n represents the number of labeled data; Generate candidate thresholds, which are dynamic thresholds, as follows: t′=t0+k*(i-1) Where t′ is the candidate threshold, t0 is the initial threshold, and t0 = min(s1,...,s n ), 1≤i≤n; The training set is divided and the model is trained using different candidate thresholds. The model performance is evaluated on the validation set, and the candidate threshold with the best comprehensive score is selected as the final splitting threshold t. The sample pairs are labeled as positive or negative samples based on the threshold t.

4. The generative semantic fusion recommendation method for service metadata as described in claim 3, characterized in that, The operation of mining difficult negative samples in S103 is as follows: for each metadata x in the training set... i Calculate the cosine approximation of the metadata in the recall set with all metadata in the training set; i Arranged in descending order of approximation; Metadata with similarity values ​​within a preset threshold range is selected to form a dataset x. i The difficult negative sample set.

5. The generative semantic fusion recommendation method for service metadata as described in claim 4, characterized in that, S2 includes: S201: Based on the semantic augmentation dataset, the pre-trained Embedding model is fine-tuned with all parameters using the domain category labels of metadata to perform multi-class classification tasks; S202: Based on the high-order metadata training set, the contrastive learning loss function is used to further train the fine-tuned Embedding model; S203: On the validation set, through adaptive iterative optimization, the optimal parameter combination is selected, and the difficult negative sample mining and embedding model training process is iteratively executed until the termination condition is met, thus obtaining a domain adaptive semantic understanding model.

6. The generative semantic fusion recommendation method for service metadata as described in claim 5, characterized in that, The full parameter fine-tuning operation in S201 includes: Based on the augmented dataset D′, a classification head is added and fine-tuned using the following multi-class cross-entropy loss function: Where C represents the total number of categories, y i p represents the one-hot encoding of the real label. i This represents the Softmax probability output by the model.

7. The generative semantic fusion recommendation method for service-oriented metadata as described in claim 5 or 6, characterized in that, The contrastive learning loss function in S202 is the InfoNCE loss function, and its expression is as follows: Among them, h i It is the embedding representation of the initial metadata. It is the embedding representation of positive metadata. It is the embedding representation of negative metadata, sim(h) i ,h j ) is the approximation calculation between two embeddings, and τ is the temperature hyperparameter.

8. The generative semantic fusion recommendation method for service metadata as described in claim 7, characterized in that, The adaptive iterative optimization of S203 includes: testing different threshold combinations on the validation set through grid search to evaluate macro and micro indicators; setting an early stopping mechanism to terminate the iteration; and finally retraining the model with the full dataset.

9. The generative semantic fusion recommendation method for service metadata as described in claim 8, characterized in that, S3 includes: S301: Employs a domain-adaptive semantic understanding model to encode query metadata into vector representations and efficiently retrieves a coarse-ranked candidate metadata set, ten times the size of the predetermined recommended metadata, from the recall set using an approximate nearest neighbor search algorithm. S302: Construct prompt words containing the coarse-ranked candidate metadata information, domain knowledge, and specific task instructions, and input them into the Large Language Model (LLM); S303: Using the large language model LLM, perform deep semantic analysis and relevance judgment on the coarsely ranked candidate metadata, and output the final recommended metadata after filtering and reordering.

10. The generative semantic fusion recommendation method for service metadata as described in claim 9, characterized in that, In step S303, the final recommendation metadata of the Large Language Model (LLM) is in a structured format. The final recommendation result is obtained by parsing the final recommendation metadata using regular expressions.

Images