A cold operation sorting method based on large model semantic enhancement

Abstract

The application discloses a cold logistics sorting method based on large model semantic enhancement, relates to the technical field of cold logistics freight logistics platform design, and comprises the following steps: based on a cold logistics freight source scene, performing RAG text vectorization design, extracting key fields, generating a special prompt word template, combining scene design Instruct, guiding a large model to obtain key vector dimensions used for freight source similarity judgment; after performing vector distillation compression processing on the key vector dimensions, taking the compressed low-dimensional vectors as new features, training and predicting through a sorting model to perform cold logistics sorting. Through the large model semantic enhancement technology, the application integrates text semantic information into the feature system of the sorting model, solves the fusion difficulty between the large model vector and the traditional sorting features through distillation compression, and provides a precise and efficient technical solution for intelligent matching of the cold logistics freight logistics platform.

Description

Technical Field

[0001] This invention relates to the field of cold chain freight logistics platform design technology, and more specifically, to a cold chain sorting method based on large model semantic enhancement. Background Technology

[0002] In the field of cold chain freight logistics platform design technology, for freight logistics cargo recommendation and ranking scenarios, the ranking model aims to accurately prioritize candidate cargo and push highly matched cargo to drivers, thereby improving the platform's delivery efficiency and driver-cargo conversion rate. Currently, the mainstream deep learning ranking models in the industry, such as DeepFM, DIN, and DCN, all rely heavily on numerical and categorical structured features such as driver historical behavior, cargo attributes, and driver-cargo cross-correlation to complete the matching and ranking. However, existing technologies have many core defects and are difficult to meet the high-precision business requirements.

[0003] In actual freight operations, cargo listings posted by cargo owners contain a large amount of high-value unstructured text information, such as cargo names, transportation temperature control, loading and unloading requirements, and emergency remarks. This information contains rich implicit business constraints and semantic matching information. However, traditional ranking models can only simply encode or discard this type of text, failing to mine deep semantic features. This results in a single dimension of cargo feature characterization and low information utilization, which greatly limits the accuracy of ranking and matching. Although large models have powerful text semantic understanding capabilities and can effectively mine refined features of cargo text, directly applying them to online ranking systems presents significant technical bottlenecks. Not only are real-time inference latency high and computational costs high, making it impossible to meet the high-concurrency, low-latency operation requirements of the business, but the high-dimensional semantic vectors of large models also differ significantly from traditional structured feature spaces. Simple splicing can easily lead to feature space chaos and model convergence difficulties. The industry has always lacked an efficient and standardized fusion solution.

[0004] Meanwhile, existing ranking schemes rely heavily on manual design for feature engineering, resulting in long iteration cycles and poor scenario adaptability. Manually generated features can only capture surface statistical patterns and cannot characterize complex semantic relationships and implicit constraints in text, thus limiting their feature representation capabilities. Furthermore, traditional technologies suffer from a severe cold start problem for new listings. Newly listed listings lack historical structured features and can only be supplemented with default values, leading to extremely poor feature characterization capabilities and consequently low accuracy in ranking and exposure of new listings, as well as poor conversion rates.

[0005] In summary, existing freight source ranking technologies suffer from four core problems: insufficient characterization of freight source features, inefficiency in enabling ranking through the semantic capabilities of large models, low degree of automation in feature engineering, and poor cold start performance for new freight sources. These limitations restrict model accuracy and scenario adaptability. Therefore, there is an urgent need for a new freight source ranking technology that integrates the semantic enhancement capabilities of large models and achieves end-to-end transformation of text semantic features. Summary of the Invention

[0006] To address the aforementioned issues, the present invention aims to provide a cold chain freight ranking method based on large model semantic enhancement, which seeks to overcome the technical shortcomings of existing cold chain freight ranking technologies, such as insufficient characterization of freight features, difficulty in efficiently integrating the semantic capabilities of large models, strong reliance on manual feature engineering, and poor cold start performance for new freight sources.

[0007] To achieve the above technical objectives, this application provides a cold chain sorting method based on large model semantic enhancement, comprising the following steps: Based on the cold chain logistics cargo scenario, we performed RAG text vectorization design, extracted key fields, and generated dedicated prompt word templates to combine with scenario design instructions to guide the large model to obtain key vector dimensions for cargo similarity judgment. After performing vector distillation compression on the key vector dimensions, the compressed low-dimensional vector is used as a new feature and trained and predicted by the ranking model to perform cold chain ranking.

[0008] Preferably, when performing RAG text vectorization design, all business fields of a single source of goods are extracted, and all fields are concatenated into standardized natural language query text according to a fixed standard format; The standardized natural language query text is input into the large model, and a unique semantic vector for the source of goods is generated based on the description of the source of goods, so as to complete the RAG vector generation.

[0009] Preferably, when generating a dedicated prompt word template, the semantic weight priority of each information dimension is set to guide the large model to focus on the core key information of freight matching when encoding text, and weaken the secondary vehicle specification information, so as to generate a dedicated prompt word template.

[0010] Preferably, when setting semantic weight priority, the semantic weight priority is set according to the matching degree of origin and destination, cargo source remarks, vehicle resource matching degree, and secondary cargo source type in sequence.

[0011] Preferably, when performing vector distillation compression design, a neural network distiller is constructed to compress the original high-dimensional vector into a compressed vector with significantly reduced dimensionality, while retaining key semantic information. The neural network distiller is a fully connected neural network.

[0012] Preferably, the fully connected neural network includes a shared underlying encoder and a task-specific decoder, wherein the shared underlying encoder consists of an input layer and three hidden layers connected in sequence; Each hidden layer consists of a fully connected layer + ReLU activation + Batch Norm, and dimensionality reduction is performed sequentially. The task-specific decoder includes an output layer that is a fully connected layer.

[0013] Preferably, when performing vector distillation compression design, the total loss function is obtained by weighting and summing the cosine similarity loss, mean square error loss, contrastive learning loss, and ranking task adaptation loss, and then used for distillation training.

[0014] Preferably, in response to a cold start of a new source of goods, a predefined default vector is used as a substitute when the RAG vector has not yet been generated. The default vector includes the global mean vector, the category mean vector, and the zero vector.

[0015] Preferably, after a new source of goods is released, an asynchronous task is automatically triggered to call the large model to generate semantic vectors, the new source of goods is put into a high-priority queue, and batch vector generation is carried out during off-peak periods; after the vector generation is completed, the default value is automatically replaced; at the same time, the pre-trained knowledge of the large model is used to quickly encode the new source of goods.

[0016] Preferably, when generating vectors, an offline batch pre-generation method is used for vector generation.

[0017] Based on the same inventive concept, this invention also discloses a cold chain sorting system based on large-model semantic enhancement, comprising: The vector design module is used to perform RAG text vectorization design based on cold chain logistics cargo scenarios, extract key fields, and generate dedicated prompt word templates. Combined with scenario design instructions, it guides large models to obtain key vector dimensions for cargo similarity judgment. The sorting module performs vector distillation compression on the key vector dimensions, then uses the compressed low-dimensional vector as a new feature, and trains and predicts it through the sorting model to perform cold chain sorting.

[0018] The present invention discloses the following technical effects: This invention integrates textual semantic information into the feature system of a ranking model through large-model semantic enhancement technology, and solves the fusion problem between large-model vectors and traditional ranking features through distillation compression, providing a precise and efficient technical solution for intelligent matching of cold chain freight logistics platforms. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the method described in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] like Figure 1 As shown, this invention provides a cold chain ranking method based on large model semantic enhancement. By introducing large model semantic enhancement technology and combining it with the core information of the cold chain logistics scenario, unstructured text information on the source side, such as cargo name, category, origin and destination, and remarks, is transformed into semantic vectors. The generated vectors are then distilled and compressed, and the compressed low-dimensional vectors are used as additional feature inputs to the ranking model. Existing ranking models are then used for training and prediction, ultimately improving the ranking performance.

[0023] In one embodiment, the core architecture of the present invention is divided into three main stages, forming a complete text semantic enhancement ranking chain: The first phase involves RAG text vectorization design. For cold chain logistics cargo scenarios, a dedicated prompt template is designed, including but not limited to key fields such as cargo name, category, detailed origin and destination addresses, remarks, price, vehicle length, and required vehicle type. This is combined with scenario-based instruction design to guide the large-scale model, informing it of the key vector dimensions for cargo similarity judgment. This allows the large-scale model to fully learn and guide the direction of vector characterization. The large-scale model performs semantic understanding on these structured cargo texts, generating cargo-level semantic vector representations. Each cargo source generates a fixed-length vector that encodes multi-dimensional semantic information, effectively integrating unstructured text information neglected in traditional methods into the ranking system.

[0024] In one implementation, RAG text vectorization is designed.

[0025] For example, the Prompt design for generating the source vector is as follows: f"Instruct: Given the description of the goods, generate a goods vector and retrieve similar goods." # The priority for calculating cargo source vectors is as follows: origin city, destination city, cargo source remarks, required vehicle length, and required vehicle type. # f"Query: Cargo Name: {row['cargo_name']}, Weight: {row['cargo_weight']}, " \ # f"Volume: {row['cargo_capacity']} cubic meters, Required vehicle length: {str(row['truck_len_list'])}, Required vehicle type: {str(row['truck_type_list'])}," \ # f"Starting location: {row['load_addr']}, " \ # f"Destination: {row['unloading_addr']}, Source Notes: {row['description']}, Secondary Source Type: {row['cargo_type_2']}"

[0026] Final assembled Prompt example: Instruct: Given cargo description information, generate a cargo vector and retrieve similar cargo. The cargo vector calculation priority is as follows: starting point city, destination city, cargo remarks, required vehicle length, required vehicle type. Query: cargo name: [cargo name], weight: [weight], volume: [volume] cubic meters, required vehicle length: [vehicle length list], required vehicle type: [vehicle type list], starting point: [starting address], destination: [destination address], cargo remarks: [remarks], secondary cargo type: [secondary cargo type].

[0027] For example, firstly, all business fields of a single cargo source in the database are read, covering cargo name, cargo weight, cargo volume, required vehicle length, required vehicle type, loading start address, unloading destination address, cargo source remarks, and secondary cargo sub-type. Then, all fields are concatenated into standardized natural language query text according to a fixed format. Subsequently, a unified execution instruction is issued to the large model, specifying that the task is to generate a unique cargo source semantic vector based on the cargo source description to complete the RAG vector generation step, and to retrieve similar cargo sources with similar business based on this vector. At the same time, the semantic weight priority of each information dimension is specified for the large model when calculating the vector, from high to low: loading start city, unloading destination city, cargo source remarks, required vehicle length, and required vehicle type. This guides the model to focus on the core key information of freight matching when encoding text, while weakening secondary vehicle specification information, in order to generate a dedicated prompt word template.

[0028] For example, after receiving standardized cargo query text, the large model performs semantic encoding based on given weight constraint rules, outputs a high-dimensional semantic vector of the corresponding cargo, and completes the RAG vector generation stage. The generated cargo semantic vector is then processed by vector distillation and compression to transform it into low-dimensional lightweight features, which are then input into the ranking model along with the original discrete and continuous features of the cargo. Subsequently, model training, online deployment, and real-time prediction and inference are carried out in sequence, ultimately realizing intelligent ranking of cold chain cargo that integrates text semantic information. At the same time, relying on the large model's semantic parsing capability of cargo text, the feature characterization of new cargo can be completed without relying on historical statistical features of cargo, thus solving the cold start problem of new cargo.

[0029] In one embodiment, regarding the Prompt design: The Prompt design of this invention closely integrates with the business characteristics of cold chain cargo and is designed according to the following priorities: (1) Origin city and destination city: These are the highest priority factors because the core value of cold chain logistics lies in point-to-point transportation services. The degree of matching between the origin and destination directly determines the feasibility and economy of the transportation route and is the primary condition for judging whether the source of goods is worth recommending.

[0030] (2) Cargo source notes: include important temperature control requirements, loading and unloading conditions, timeliness requirements, etc., such as the need for full freezing, urgent departure of goods, etc. These information directly affect whether the driver can meet the requirements of the cargo owner.

[0031] (3) Required vehicle length and required vehicle type: These determine the matching degree of vehicle resources. Different sources of goods may require different specifications of vehicles, such as refrigerated trucks and dual-temperature zone trucks.

[0032] (4) Secondary source type: reflects the refined classification of sources of goods, which helps to distinguish the subtle differences between similar sources of goods.

[0033] For example, this prioritization design allows large models to focus more on the feature dimensions that have the greatest impact on the ranking results when generating vectors, achieving a more accurate characterization of the source of goods, thereby improving the overall performance of the ranking model.

[0034] The second stage involves vector distillation compression. Semantic vectors generated by large models are typically high-dimensional, such as 1024 or 4096 dimensions. Directly concatenating these vectors into the features of a ranking model can lead to problems such as feature dimensionality explosion, excessive computational burden, and overfitting. This invention introduces knowledge distillation technology. By constructing a neural network distiller, it takes the original high-dimensional vector as input and learns to compress it into a compressed vector with significantly reduced dimensionality, such as 32 dimensions, while retaining key semantic information.

[0035] In one embodiment, the design is for vector distillation compression.

[0036] In one embodiment, for the design of the distillation model architecture, the present invention uses a fully connected neural network as a distiller to compress high-dimensional semantic vectors, such as an input dimension of 4096, into low-dimensional feature vectors, such as 32-dimensional output.

[0037] For example, the distillation model architecture is as follows: (1) Shared Encoder: Input layer: 4096-dimensional original semantic vector; Hidden layer 1: 4096→1024, fully connected layer + ReLU activation + Batch Norm; Hidden layer 2: 1024→256, fully connected layer + ReLU activation + Batch Norm; Hidden layer 3: 256→64, fully connected layer + ReLU activation + Batch Norm; (2) Task-specific decoders: Output layer: 64→32, fully connected layer (compressed to the dimension required for sorting).

[0038] For example, the distillation model structure is described as follows: a fully connected (FC) network is used as the distiller, and the progressive dimensionality reduction (4096→1024→256→64→32) helps to gradually extract key semantic information; Batch Norm is added to each layer to accelerate convergence and improve generalization ability, and nonlinear transformation capability is introduced through the ReLU activation function.

[0039] For example, in designing the distillation loss function, this invention uses a combination of multiple loss functions for distillation training to ensure that the compressed vector retains the semantic information of the original vector while also having characteristics suitable for sorting tasks.

[0040] (1) Cosine Similarity Loss: L_cos=1-cos(Teacher_vec,Student_vec) Here, cos represents cosine similarity, Teacher_vec is the original vector, and Student_vec is the compressed vector. This loss function ensures that the compressed vector and the original vector maintain the same direction, preserving semantic similarity.

[0041] (2) Mean Squared Error Loss (MSE Loss): L_mse=||Teacher_vec-Student_vec||² The loss function ensures that the compressed vector is numerically close to the original vector while preserving the absolute semantic information of the vector.

[0042] (3) Contrastive Loss: L_contrast=-log(exp(sim(pos) / τ) / Σexp(sim(neg) / τ)) Where sim represents vector similarity, τ is the temperature parameter, pos is a positive sample pair, and neg is a negative sample pair. This loss function enhances the discriminative power of vectors by contrastively learning to bring vectors of similar products closer together and widen vectors of dissimilar products apart.

[0043] (4) Ranking-awareLoss: L_rank=Σmargin(pred_i-pred_j) Where pred is the predicted score of the ranking model, and margin is the ranking interval. This loss function ensures that the compressed vector has good ranking performance in downstream ranking tasks.

[0044] Comprehensive loss function: L_total=λ1·L_cos+λ2·L_mse+λ3·L_contrast+λ4·L_rank Wherein, λ1, λ2, λ3, and λ4 are weight hyperparameters, and it is recommended to take values ​​of λ1=0.3, λ2=0.3, λ3=0.2, and λ4=0.2.

[0045] For example, through the analysis of the advantages of using the distilled model for ranking models, the compressed vector after distillation can be better used for ranking models, mainly for the following reasons: (1) Dimensionality adaptability: Ranking models usually use low-dimensional feature vectors (such as 32-128 dimensions). Directly using a 4096-dimensional high-dimensional vector will lead to dimensionality imbalance and affect model learning. The distilled 32-dimensional vector is more compatible with the other feature dimensions of the ranking model, which facilitates model fusion and learning.

[0046] (2) Noise filtering capability: The distillation process is essentially an information extraction process. Fully connected networks can learn which semantic information is important for the ranking task and which is noise. Through training, the distiller can automatically filter out high-frequency noise and redundant information that is not helpful for ranking, while retaining the core semantic features.

[0047] (3) Semantic space alignment: During training, the distiller learns the mapping relationship from the original semantic space to the sorting feature space. The compressed vector has been "sorted" in the semantic space and is in the same semantic space as other features of the sorting model, so it can work together better.

[0048] (4) Avoiding the curse of dimensionality: Directly concatenating high-dimensional vectors can lead to an explosion of feature dimensions, causing problems such as increased sparsity and increased risk of overfitting. Distilled low-dimensional vectors effectively avoid these problems, enabling the ranking model to be trained stably.

[0049] (5) Maintaining semantic distinguishability: By designing the contrastive loss function, the distiller maintains the semantic distinguishability of the original vectors during the compression process. Similar sources remain similar after compression, and dissimilar sources remain different, ensuring the semantic effectiveness of the ranking features.

[0050] (6) Improved computational efficiency: Compared with 4096-dimensional vectors, 32-dimensional vectors have significant computational efficiency advantages in feature processing and matrix operations of sorting models, and can meet the low latency requirements of sorting systems.

[0051] The third stage involves designing a ranking model that enhances RAG semantic features using compressed low-dimensional vectors as new features. These vectors are then integrated with the feature systems of existing ranking models such as DeepFM, DIN, and PLE, and used for training and prediction. This design allows RAG semantic features to be smoothly integrated into existing ranking systems without requiring significant adjustments to the existing model architecture.

[0052] In one embodiment, the design of a solution to the cold start problem for new product sources addresses the issue that newly listed products, due to the lack of RAG semantic vectors, will experience a cold start. This invention employs the following multi-layered processing scheme: 1. Default Vector Substitution Strategy: When RAG vectors have not yet been generated, predefined default vectors are used for substitution. Specific methods include: Global mean vector: Uses the mean of all historical source vectors as the default value; Category average vector: Based on the secondary category of the goods, the average of the historical goods vectors of this category is used as the default value; Zero vector: For sources of goods whose category cannot be determined, use an all-zero vector as the most conservative default value.

[0053] 2. Asynchronous vector generation strategy: After a new product is released, an asynchronous task is automatically triggered to call the large model to generate semantic vectors. Priority queue: New supplies enter the high-priority queue to ensure rapid generation; Batch processing: Generate vectors in batches during off-peak hours to reduce the cost of calling large models; Incremental update: After the vectors are generated, the default values ​​are automatically replaced, without the need to retrain the sorting model; For example, fast encoding based on prior knowledge of a large model utilizes the pre-trained knowledge of the large model to quickly encode new sources of goods: Category Feature Extraction: Extract semantic information at the category level based on the secondary category information of the goods. Text summary generation: Generates detailed text information about the source of goods. Feature vector generation: Based on the text generated in the first two steps, large model vectorization and vector distillation are performed. Cold start processing flow: New product release → Check if there is a RAG vector → No → Use default vector / category average vector → Simultaneously trigger asynchronous vector generation → Automatically replace the default value after vector generation is completed → Apply the sorting model; Through the above multi-level processing scheme, this invention can effectively solve the cold start problem of new products, ensuring that new products can obtain a reasonable sorting position even before semantic vectors are generated, and gradually improve the sorting effect of new products through the asynchronous update mechanism.

[0054] In one implementation, the design of the training and prediction process is described.

[0055] For example, considering the inference latency and cost of large models, this invention uses an offline batch pre-generation method for vector generation: Training phase: 1. In advance, call the large model to generate semantic vectors for all historical sources; 2. Compress the high-dimensional vectors into 32-dimensional low-dimensional vectors using a distillation device; 3. Store the compressed vectors as offline feature data; 4. Fuse the vector features with other features of the existing ranking model; 5. Train the ranking model using the fused features.

[0056] Prediction phase: 1. Read the pre-generated compressed vector directly based on the cargo source ID; 2. No need to call the large model in real time, ensuring low latency of the ranking system; 3. Concatenate the vector features with other features and input them into the ranking model; 4. Output the ranking results and recommend top cargo sources to drivers.

[0057] New product source processing: 1. Use the default vector or category mean vector when publishing; 2. Trigger an asynchronous task to generate RAG vectors; 3. Automatically replace the default value after vector generation; 4. Put the new vector into the ranking model for prediction to improve the ranking model's performance.

[0058] For example, this invention specifically addresses the unique business characteristics of cold chain logistics scenarios, leveraging customized prompt word templates to mine deep semantic information from unstructured text such as cold chain cargo remarks and core cargo descriptions. It also incorporates knowledge distillation compression technology to achieve lightweight transformation of semantic vectors, effectively breaking down the technical barriers between large-scale models and traditional ranking models. This enables efficient end-to-end transformation of unstructured text data into usable features for ranking models, replacing traditional manual feature design and fundamentally optimizing the intelligent matching and ranking effect of drivers and cargo in cold chain scenarios.

[0059] For example, the present invention, through the above-mentioned core technical solutions, can achieve several beneficial effects: First, it significantly improves the ability to characterize cargo features. By leveraging the deep text semantic understanding capabilities of large models, it fully mines the key constraint information implicit in the cold chain cargo text, such as transportation temperature control, loading and unloading requirements, and urgent delivery, thus overcoming the shortcomings of traditional structured features being of single dimension and lacking semantic information, and improving the accuracy of driver-cargo matching. Second, it achieves efficient end-to-end transformation of text semantic features. Through semantic vectorization combined with knowledge distillation compression technology, it compresses the high-dimensional semantic vectors output by large models into low-dimensional standardized features, adapting them to the feature space of ranking models, and solving the problems of high latency in real-time inference, high computing power cost, chaotic feature space, and model difficulty in large models. First, it addresses the convergence issue to meet the needs of low-latency, high-concurrency online business. Second, it automates the entire process of cargo feature engineering by integrating prompt word engineering, semantic vectorization, vector dimensionality reduction, and model inference. This eliminates the reliance on manual feature design, shortens the feature iteration cycle, overcomes the limitations of manual feature expression capabilities, and significantly improves feature construction efficiency and scenario adaptability. Third, it effectively solves the cold start problem for new cold chain cargo by using general prior knowledge from a large model to semantically characterize the text information of newly listed cargo. This eliminates the need to rely on historical statistical structured features, avoids feature distortion caused by default value completion, accurately characterizes the attributes and matching requirements of new cargo, and effectively improves the accuracy of new cargo ranking and exposure and the conversion rate.

[0060] In summary, this invention, by adapting to cold chain logistics operations and setting a dedicated source vector Prompt template with semantic dimension priorities, combined with RAG semantic vectorization technology, fully explores unstructured textual information of the source cargo to compensate for the deficiencies in source cargo feature characterization. It also establishes a complete "semantic vectorization—distillation compression—feature fusion" chain, utilizing a fully connected distillation network with a multi-loss function combination distillation strategy to compress a 4096-dimensional high-dimensional semantic vector to 32 dimensions. While preserving core semantics, it achieves efficient serialization of large-scale model semantic capabilities into a ranking model. Relying on standardized prompt word templates, it automatically completes text feature extraction and semantic modeling, eliminating reliance on manual feature engineering, shortening iteration cycles, and reducing labor costs. Simultaneously, it employs a multi-layered cold-start processing scheme that integrates large-scale model prior knowledge, category mean vectors, and asynchronous generation mechanisms to optimize the ranking effect of new sources of cargo. Overall, it achieves end-to-end fusion of large-scale model semantic enhancement and ranking model, fundamentally solving the four major limitations of traditional ranking models: weak source cargo feature characterization, difficulty in implementing large-scale model capabilities, low automation of feature engineering, and poor cold-start effect for new sources of cargo. This comprehensively improves the accuracy of cold chain cargo matching and ranking, as well as the efficiency of online business operations.

[0061] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A cold chain sorting method based on large model semantic enhancement, characterized in that, Includes the following steps: Based on the cold chain logistics cargo scenario, we implemented RAG text vectorization design, including extracting key fields and generating dedicated prompt word templates. Combined with scenario design instructions, this guides the large model to obtain key vector dimensions for cargo similarity judgment and generate RAG vectors. After performing vector distillation compression on the key vector, the compressed low-dimensional vector is used as a new feature and trained and predicted by the ranking model to perform cold chain ranking. When generating the dedicated prompt word template, the semantic weight priority of each information dimension is set to guide the large model to focus on the core key information of freight matching when encoding text, and weaken the secondary vehicle specification information, so as to generate the dedicated prompt word template. When setting semantic weight priority, the semantic weight priority is set according to the matching degree of origin and destination, cargo source remarks, vehicle resource matching degree, and secondary cargo source type in sequence. When performing vector distillation compression design, a neural network distiller is constructed to compress the original high-dimensional vector into a compressed vector with significantly reduced dimensionality, while retaining key semantic information. The neural network distiller is a fully connected neural network. The fully connected neural network includes a shared underlying encoder and a task-specific decoder, wherein the shared underlying encoder consists of an input layer and three hidden layers connected in sequence; Each of the hidden layers consists of a fully connected layer + ReLU activation + Batch Norm, and dimensionality reduction processing is performed sequentially. The task-specific decoder includes a fully connected layer as the output layer; When performing vector distillation compression design, cosine similarity loss, mean square error loss, contrastive learning loss, and ranking task adaptation loss are used to obtain the total loss function through weighted summation for distillation training. In response to the cold start of new sources of goods, when the RAG vector has not yet been generated, a predefined default vector is used as a substitute, wherein the default vector includes at least one of the global mean vector, the category mean vector, and the zero vector; After a new source of goods is released, an asynchronous task is automatically triggered to call the large model to generate semantic vectors. The new source of goods is added to a high-priority queue, and batch vectors are generated during off-peak periods. After the vectors are generated, the default values ​​are automatically replaced. At the same time, the pre-trained knowledge of the large model is used to quickly encode the new source of goods.

2. The cold chain sorting method based on large model semantic enhancement according to claim 1, characterized in that: When implementing RAG text vectorization design, all business fields of a single source of goods are extracted, and all fields are concatenated into standardized natural language query text according to a fixed standard format. The standardized natural language query text is input into the large model, and a unique semantic vector for the source of goods is generated based on the description of the source of goods, so as to complete the RAG vector generation.

3. The cold chain sorting method based on large model semantic enhancement according to claim 1, characterized in that: When generating vectors, an offline batch pre-generation method is used.

Images