Large model-based sequence generation scoring method, device, equipment and medium

Abstract

This application discloses a sequence generation and scoring method, apparatus, device, and medium based on a large model, relating to the field of artificial intelligence data processing technology. The method includes: mapping nodes and fusing features of multi-source heterogeneous data of a target object using a preset knowledge graph to obtain object subgraph features; converting the object subgraph features into soft cue vectors using a preset feature projection layer, and concatenating them with a preset task cue template to form a fused cue sequence; inputting the fused cue sequence into a pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences; inputting each candidate sequence into a scoring model to extract temporal features and outputting a comprehensive score value; selecting a target task processing sequence from the multiple candidate task processing sequences based on the comprehensive score value, and sending it to the target business system. This application can improve the alignment effect between multi-source heterogeneous data and the input space of a pre-trained large model, and improve the generation quality and scoring accuracy of candidate task processing sequences.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence data processing technology, and in particular to sequence generation scoring methods, apparatus, devices and media based on large models. Background Technology

[0002] With the development of information systems, various business scenarios have accumulated a large amount of multi-source heterogeneous data. For task processing workflows with long sequence dependencies, traditional manual processing methods usually rely on empirical rules, making it difficult to fully utilize the relational features in structured data, unstructured text, and knowledge graphs. To improve the processing efficiency of multi-source heterogeneous data and enhance the generation quality of candidate task processing sequences, utilizing artificial intelligence technologies for feature fusion, sequence generation, and sequence scoring and filtering of heterogeneous data has become an important requirement in the field of intelligent data processing.

[0003] Currently, sequence generation data processing technologies are mainly divided into two categories: one is the traditional expert system based on established rules or decision trees, which mechanically matches input data through preset hard logic; the other is the generative system that introduces a large language model, which usually directly concatenates unstructured text and inputs it into a general large model, relying on the model's own pre-trained corpus to directly generate candidate task processing sequences.

[0004] However, the above approach has significant limitations in practical applications. On the one hand, traditional rule systems lack flexibility and struggle to effectively handle personalized features hidden within complex heterogeneous data. On the other hand, directly using generative pre-trained large models to output long sequences can easily lead to illusions, as the generated sequences cannot guarantee consistent quality and constraint satisfaction during long-term task processing. Therefore, how to generate target task processing sequences that balance sequence quality indicators and constraints based on multi-source heterogeneous data has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a sequence generation and scoring method, apparatus, device and medium based on a large model, which aims to solve the technical problem of how to generate target task processing sequences based on multi-source heterogeneous data that take into account both sequence quality indicators and constraints.

[0006] To achieve the above objectives, this application proposes a sequence generation scoring method based on a large model, the method being executed by an electronic device, the method comprising: Acquire multi-source heterogeneous data of target objects collected or stored by the target business system, perform node mapping and feature fusion on the multi-source heterogeneous data through a preset knowledge graph, and aggregate node features through a graph neural network to obtain object subgraph features; The object subgraph features are transformed into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and then concatenated with the text embedding vectors corresponding to the preset task cue template to obtain a fused cue sequence. The fusion prompt sequence is input into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences; Each of the candidate task processing sequences is input into a pre-trained scoring model, and the temporal features of each candidate task processing sequence are extracted through the scoring model. Based on the aforementioned temporal characteristics, the scoring model outputs a comprehensive score value for each candidate task processing sequence that satisfies the preset sequence quality index and preset constraints. The target task processing sequence is selected from multiple candidate task processing sequences based on the comprehensive score, and then sent to the target business system through a communication interface.

[0007] Furthermore, to achieve the above objectives, this application also proposes a sequence generation and scoring device based on a large model, the device comprising: The mapping module is used to acquire multi-source heterogeneous data of target objects collected or stored by the target business system, perform node mapping and feature fusion on the multi-source heterogeneous data through a preset knowledge graph, and aggregate node features through a graph neural network to obtain object subgraph features. The concatenation module is used to transform the features of the object subgraph into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and concatenate them with the text embedding vectors corresponding to the preset task cue template to obtain a fused cue sequence. The decoding module is used to input the fused prompt sequence into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences; The feature extraction module is used to input each of the candidate task processing sequences into a pre-trained scoring model, and extract the temporal features of each of the candidate task processing sequences through the scoring model. The model output module is used to output a comprehensive score value for each candidate task processing sequence that meets the preset sequence quality index and preset constraints based on the temporal features and through the scoring model. The filtering and sending module is used to filter out the target task processing sequence from multiple candidate task processing sequences based on the comprehensive score value, and send the target task processing sequence to the target business system through a communication interface.

[0008] Furthermore, to achieve the above objectives, this application also proposes a sequence generation and scoring device based on a large model, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the sequence generation and scoring method based on a large model as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the sequence generation and scoring method based on a large model as described above.

[0010] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the sequence generation and scoring method based on a large model as described above.

[0011] One or more technical solutions proposed in this application have at least the following technical effects: First, a pre-defined knowledge graph is used to process multi-source heterogeneous data of the target object, mapping scattered information to graph nodes and fusing them into object subgraph features, thus achieving a structured relational representation of complex heterogeneous data. Next, a pre-defined feature projection layer converts the object subgraph features into soft cue vectors, which are then concatenated with a pre-defined task cue template to form a fused cue sequence, achieving cross-modal alignment between graph network features and natural language text features in the input space of the pre-trained large model. Then, the fused cue sequence is input into the pre-trained large model for autoregressive decoding, generating multiple candidate task processing sequences, thereby leveraging the large model's contextual modeling capabilities to explore diverse sequence paths. Subsequently, each candidate task processing sequence is input into a pre-trained scoring model to extract temporal features reflecting the sequential order of task steps, capturing dynamic dependency patterns in long-sequence task processing. Based on this, the scoring model calculates a comprehensive score for each candidate task processing sequence under pre-defined sequence quality indicators and constraints, achieving an objective quantitative evaluation of the candidate task processing sequences. Finally, the target task processing sequence is selected from the candidate task processing sequences based on the comprehensive score and output to the target business system. This application can generate target task processing sequences based on multi-source heterogeneous data, taking into account both sequence quality indicators and constraints, thereby improving the sequence generation quality, scoring and filtering accuracy, and processing efficiency of complex heterogeneous data. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the first embodiment of the sequence generation and scoring method based on a large model in this application. Figure 2 This is a flowchart illustrating the second embodiment of the sequence generation and scoring method based on a large model in this application. Figure 3 This is a schematic diagram of the module structure of the sequence generation and scoring device based on a large model according to an embodiment of this application; Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the sequence generation and scoring method based on a large model in the embodiments of this application.

[0015] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0017] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0018] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or sequence generation scoring system capable of realizing the above functions. The following description uses an electronic device as an example to illustrate this embodiment and the subsequent embodiments.

[0019] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0020] This application processes multi-source heterogeneous data, which is actual object state data collected or stored by the target business system, including at least structured attribute data and unstructured text data. Through the collaborative processing of a pre-set knowledge graph, graph neural network, feature projection layer, pre-trained large model, and scoring model, data from different sources, modalities, and dimensions are transformed into fused prompt sequences that can be used as input to the pre-trained large model. Temporal feature extraction and constraint scoring are then performed on the generated candidate task processing sequences. This application is not simply a mathematical model building method, nor is it merely a rule calculation for abstract data. Rather, it is a multi-source heterogeneous data processing method based on electronic devices, used to solve the problems of mismatch between heterogeneous data and the input space of the pre-trained large model, and the lack of constraint filtering in long sequence generation results. Through the above processing, the number of invalid candidate task processing sequences entering the target business system can be reduced, the overhead of repeated sequence transmission and storage can be reduced, and the execution efficiency of the candidate sequence generation and filtering process can be improved.

[0021] Based on this, the first embodiment of this application provides a sequence generation scoring method based on a large model, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the sequence generation and scoring method based on a large model in this application.

[0022] In this embodiment, the sequence generation and scoring method based on a large model is executed by an electronic device, and the method includes steps S10 to S60: Step S10: Obtain multi-source heterogeneous data of the target object collected or stored by the target business system, perform node mapping and feature fusion on the multi-source heterogeneous data through a preset knowledge graph, and perform node feature aggregation through a graph neural network to obtain object subgraph features; It should be noted that the target business system refers to the information system that collects, stores, or receives data related to the target object and the target task processing sequence. The pre-defined knowledge graph can be network structure data containing entities and relationships between entities in the target business scenario. The target object can be an entity in the target business system that needs to generate a task processing sequence, such as a device, user, work order, data record set, or event to be processed. Multi-source heterogeneous data refers to data in different formats corresponding to the target object, such as unstructured records in text form, structured attribute data in numerical form, log data, status monitoring data, or business process data.

[0023] Multi-source heterogeneous data refers to actual object state data collected or stored by the target business system, not abstract mathematical samples. Object subgraph features refer to the local network topology numerical representation extracted from the complete knowledge graph that is related to the current state of the target object. Node mapping refers to mapping entities, events, attributes, or textual semantic information from multi-source heterogeneous data to graph nodes in a predefined knowledge graph. Feature fusion refers to combining the semantic features corresponding to structured attribute data, unstructured text data, and graph node features. Node feature aggregation refers to weighted aggregation of features corresponding to target mapped nodes, adjacent nodes, and multi-hop relationship edges using a graph neural network.

[0024] The process involves separating structured attribute data and unstructured text data from multi-source heterogeneous data. A pre-defined entity extraction model is used to extract target entity words from the unstructured text data. A pre-defined text encoder transforms these target entity words into semantic feature vectors. The cosine similarity between these semantic feature vectors and nodes in a pre-defined knowledge graph is calculated. Nodes with a cosine similarity greater than 0.85 are selected as target mapping nodes. Adjacent nodes and multi-hop edges of these target mapping nodes are extracted to construct a local graph structure. A graph neural network is then used to weighted aggregate the features of each node in the local graph structure, outputting dimensionality-reduced object subgraph features. This step transforms scattered, multi-modal real-world object data into graph features with structured topological relationships, preserving the semantic and graph structural relationships between target entities. Target entity words refer to keywords or phrases identified from unstructured text data that are related to the target object's state, task type, business event, or processing action.

[0025] Step S20: The object subgraph features are transformed into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and then concatenated with the text embedding vectors corresponding to the preset task cue template to obtain a fused cue sequence. It should be noted that the preset feature projection layer can be a multilayer perceptron or a linear transformation matrix, mainly used to adjust the feature vector dimension between different networks. The input space dimension refers to the underlying word embedding dimension or hidden layer input dimension corresponding to the input vector received by the pre-trained large model. Soft cue vectors refer to continuous vector representations that can be used as input embeddings in the pre-trained large model for computation. Preset task cue templates refer to pre-defined text templates used to constrain the model's output format, task type, or sequence generation direction. The fused cue sequence refers to the multimodal model input structure formed by combining graph topological features and text instruction features in the temporal dimension.

[0026] Specifically, the object subgraph features are input into a preset feature projection layer, and a multilayer perceptron is used for linear transformation to align with the internal word embedding dimension of the pre-trained large model. This outputs a soft cue vector with a uniform feature scale. Template text fragments contained in a preset task cue template are obtained, and these fragments are converted into text embedding vectors using a word embedding matrix. The soft cue vectors and text embedding vectors are then concatenated in chronological order to form a fused cue sequence. This step addresses the inconsistency between the graph network topology data and the input space dimension of the pre-trained large model, achieving the alignment and concatenation of cross-modal information in the underlying vector space.

[0027] Step S30: Input the fusion prompt sequence into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences; It should be noted that the pre-trained large model can be a Transformer-based language generation network that has been pre-trained on a large corpus and fine-tuned with task instructions. The candidate task processing sequence can be a text path containing a series of task processing nodes arranged chronologically, formed by multiple task nodes, action nodes, processing steps, or text fragments arranged in temporal order. Autoregressive decoding refers to the decoding method where the pre-trained large model uses the lexical units or hidden states output from the previous generation time step as the input conditions for the current generation time step, generating candidate text paths step by step. Autoregressive sampling refers to the sampling method where, according to the vocabulary probability distribution of each generation time step, lexical units are selected from the candidate vocabulary and gradually concatenated to form a text sequence. A generation time step refers to the sequence position where the pre-trained large model sequentially generates lexical units during the autoregressive decoding process. A lexical unit refers to the basic generation unit in the candidate vocabulary of the pre-trained large model, which can be a character, word, subword, or symbol fragment.

[0028] Understandably, the fused prompt sequence is input into the multi-head self-attention layer of the pre-trained large model to extract the contextual hidden state. The decoding prediction head then extrapolates the contextual hidden state step-by-step, outputting the probability distribution of candidate lexical units or candidate task nodes at each time step. Based on the vocabulary probability distribution, lexical selection and path expansion are performed from the candidate vocabulary of the pre-trained large model. A beam search algorithm is then used to retain candidate text paths with high confidence, generating multiple complete candidate task processing sequences. This step leverages the contextual modeling and language generation capabilities of the pre-trained large model to generate multiple coherent candidate task processing sequences based on the target object's context.

[0029] Step S40: Input each of the candidate task processing sequences into a pre-trained scoring model, and extract the temporal features of each of the candidate task processing sequences through the scoring model. It should be noted that the scoring model can be a scoring network trained based on manual ranking feedback to evaluate sequence quality, such as a fine-tuned RoBERTa network. The scoring model can include an embedding layer, a multi-head self-attention layer, and a feedforward network. The embedding layer transforms candidate task processing sequences into candidate sequence embeddings, while the multi-head self-attention layer and feedforward network extract features from the complete evaluation sequence. Temporal features refer to hidden layer vector representations that reflect the sequential dependencies between multiple consecutive task processing steps and the association with global contextual constraints.

[0030] Specifically, the fused prompt sequence is used as the object context embedding. Each candidate task processing sequence is decomposed into task processing nodes corresponding to multiple time steps, and then transformed into candidate sequence embeddings through the embedding layer of the scoring model. The object context embedding and the candidate sequence embedding are concatenated along the temporal dimension to obtain the complete evaluation sequence, which is used as the initial bottom-level vector of the scoring model. The self-attention mechanism within the scoring model is used to calculate the attention weights between the initial bottom-level vectors at different time steps. The initial bottom-level vectors are then weighted and summed according to the attention weights to output the temporal features of each candidate task processing sequence. This step captures the constraint relationships between each task node before and after in the long sequence task processing path, avoiding isolated evaluation of task nodes at a single time step.

[0031] Step S50: Based on the time series characteristics, output a comprehensive score value for each candidate task processing sequence that satisfies the preset sequence quality index and preset constraints through the scoring model; It should be noted that the preset sequence quality indicators can be the task completion rate of historical similar samples, the probability of state maintenance, or a probability threshold for a specific target indicator to reach a preset range. These indicators can be determined by the task completion results of historical samples, manual annotation results, the achievement status of the target indicator, or system feedback data. Preset constraints refer to the restrictive rules stipulated in the target business scenario, such as data format constraints, system interface constraints, parameter boundary constraints, step integrity constraints, and output security constraints. The comprehensive score is a numerical value that quantitatively reflects the overall performance of a specific task processing sequence in satisfying the preset sequence quality indicators and preset constraints.

[0032] Understandably, the process involves inputting temporal features into the sequence quality evaluation head of the scoring model, calculating the degree of matching between the temporal features and preset sequence quality indicators to obtain a sequence fitting score, and using the constraint penalty network within the scoring model to output the constraint violation probability triggered by the temporal features. When the constraint violation probability is greater than a preset constraint violation threshold, a preset low score is used as the constraint satisfaction score; when the constraint violation probability is less than or equal to the preset constraint violation threshold, a constraint satisfaction score is calculated based on the constraint violation probability, where the constraint satisfaction score is negatively correlated with the constraint violation probability. A comprehensive score is then calculated by weighting and summing the sequence fitting score and the constraint satisfaction score according to preset weighting parameters. This step, by introducing the constraint violation probability, helps reduce the probability of candidate task processing sequences that violate preset constraints being selected.

[0033] Step S60: Select the target task processing sequence from the multiple candidate task processing sequences based on the comprehensive score value, and send the target task processing sequence to the target business system through the communication interface.

[0034] It should be noted that the target task processing sequence refers to the final output sequence that ranks among the top in overall score among all generated candidate task processing sequences and meets the preset output conditions. The target business system can be a server, business terminal, information management system, or process scheduling system, which receives the target task processing sequence through an interface.

[0035] The target task processing sequence output to the target business system is used as a reference processing sequence for the target business system and does not directly replace manual review, permission verification, or actual execution processes in the target business system. The comprehensive score is only used for sorting and filtering among candidate task processing sequences and does not indicate that the target task processing sequence is necessarily applicable to all business states. The target business system can determine whether to execute the target task processing sequence based on permission rules, running status, or manual confirmation results.

[0036] Specifically, multiple candidate task processing sequences are sorted in descending order according to their comprehensive score. The sequence that ranks first is extracted as the first-order sequence, and its comprehensive score is checked against a preset backoff threshold of 0.6. If the comprehensive score of the first-order sequence is greater than or equal to the preset backoff threshold, it is designated as the target task processing sequence. If the comprehensive score of the first-order sequence is less than the preset backoff threshold, the electronic device extracts object type identifiers from multi-source heterogeneous data, matches these identifiers with rule indexes in a preset rule base, retrieves the corresponding preset rule sequence path, and uses this path as the target task processing sequence. The electronic device converts the target task processing sequence into a visual timeline chart format and sends it to the target business system for display via a communication interface. This step filters candidate task processing sequences with low comprehensive scores or those violating preset constraints for the target business system, providing a visual representation of the target task processing sequence. The first-order sequence refers to the candidate task processing sequence that ranks first after multiple candidate task processing sequences are sorted in descending order of their comprehensive score.

[0037] This embodiment effectively integrates the structured topological representation of knowledge graphs with the sequence generation capabilities of pre-trained large models, solving the problems of traditional rule-based systems struggling to utilize the association features of complex heterogeneous data and the lack of constraint filtering in generative model output sequences. Simultaneously, a scoring model is introduced into the multi-step task processing sequence generation chain to perform dual quantitative evaluation of sequence quality and constraints, improving the accuracy of candidate task processing sequence selection, output stability, and data processing efficiency of the target business system.

[0038] As an example, the step of converting the object subgraph features into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and concatenating them with the text embedding vectors corresponding to the preset task cue template to obtain a fused cue sequence includes: linearly mapping the object subgraph features through the preset feature projection layer to obtain an initial continuous vector with the same input space dimension as the pre-trained large model; normalizing the initial continuous vector to obtain a feature scale-aligned soft cue vector; extracting template text fragments from the preset task cue template and converting the template text fragments into text embedding vectors through a preset word embedding layer; and concatenating the soft cue vectors and the text embedding vectors according to the temporal dimension to obtain a fused cue sequence.

[0039] It should be noted that in this example, the preset feature projection layer can be a neural network containing multiple linear transformation structures, used to map the graph network topology representation to the underlying vector space of the pre-trained large model. The initial continuous vector refers to a dense floating-point matrix whose dimension matches the dimension of the word embeddings in the underlying layer of the pre-trained large model. The text embedding vector refers to the vector representation obtained after mapping the template text fragment to the continuous vector space through the preset word embedding layer. The preset word embedding layer can be a lookup table matrix at the bottom layer of the Transformer architecture, used to map the term indices corresponding to the template text fragment to real-number vectors. Feature scale alignment refers to normalizing the distribution of the initial continuous vectors to be consistent with or similar to the distribution of the input vectors of the pre-trained large model. The template text fragment refers to the text content in the preset task prompt template, consisting of natural language text, keywords, formatting tags, or control symbols.

[0040] Understandably, the first step is to input the features of the object subgraph into a preset feature projection layer. The weight matrix inside this layer is used to linearly map the graph node features so that the feature dimension is consistent with the input space dimension of the pre-trained large model after transformation, and the initial continuous vector is output.

[0041] The Layer Normalization algorithm is used to shift and scale the feature distribution of the initial continuous vector to stabilize the data distribution of local features and output a soft cue vector with feature scale alignment.

[0042] Next, the template text fragments containing task processing constraints and output format requirements in the preset task prompt template are obtained. The template text fragments are converted into corresponding text embedding vectors by using the preset word embedding layer lookup table. According to the temporal order of the model input, the soft prompt vector is used as the preceding object state vector and the text embedding vector is used as the following task prompt vector and concatenated to output the fused prompt sequence.

[0043] This example uses feature projection and normalization to concatenate the object subgraph features with the text embedding vectors corresponding to the template text fragments in the input space of the pre-trained large model, thereby improving the joint representation effect of the object subgraph features and task prompt information.

[0044] As an example, the step of inputting the fused prompt sequence into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences includes: inputting the fused prompt sequence into the feature extraction block of the pre-trained large model to extract the contextual hidden state; mapping the contextual hidden state step by step through the language decoding head in the pre-trained large model to output the vocabulary probability distribution at each time step; generating multiple candidate text paths by autoregressive sampling from the candidate vocabulary of the pre-trained large model based on the vocabulary probability distribution at each time step; calculating the generation confidence of each candidate text path through a beam search algorithm; arranging each candidate text path in descending order according to the generation confidence, and taking the first preset number of candidate text paths as candidate task processing sequences.

[0045] It should be noted that in this example, the feature extraction block refers to the hidden layer network component within the pre-trained large model, consisting of multiple Transformer network layers, used to capture input contextual dependencies. The language decoding head can be a classifier network containing a fully connected feedforward network and a Softmax normalization function, used to map high-dimensional hidden states to probability values ​​in the vocabulary space. Generation confidence refers to the cumulative log-likelihood score that quantitatively evaluates the reasonableness of a single complete text path under the language model's generation probability system. The preset number refers to the upper limit parameter of candidate branches retained at each time step during beam search. To balance the limitations of search space and computational resource overhead, its value is usually set between 3 and 10. In this example, 5 branches are specifically used to balance generation diversity and inference efficiency. The vocabulary probability distribution refers to the probability set output by the pre-trained large model for each word in the candidate vocabulary at each generation time step. The candidate vocabulary refers to the set of words that the pre-trained large model can select when generating text. The contextual hidden state refers to the hidden layer vector representation obtained by the pre-trained large model after feature extraction of the fused cue sequence. Candidate text paths refer to the text sequence paths generated step-by-step during the autoregressive sampling process of the pre-trained large model; after the candidate text paths are sorted by generated confidence and a preset number is retained, they are used as candidate task processing sequences.

[0046] Specifically, firstly, the fused cue sequence is input into the feature extraction block inside the pre-trained large model, and the contextual association between each feature vector in the fused cue sequence is extracted using a multi-head attention mechanism, and the contextual hidden state is output.

[0047] The language decoding head in the pre-trained large model is used to perform a time-step linear mapping transformation on the contextual hidden state. At each generation time step, the high-dimensional continuous hidden state vector is converted into a numerical vector with the same dimension as the candidate vocabulary of the pre-trained large model, and the occurrence probability of each word in the vocabulary at each time step is output, resulting in a vocabulary probability distribution. Based on the vocabulary probability distribution output at each time step, the top K words with the highest vocabulary probability are selected from the candidate vocabulary of the pre-trained large model, or words with a vocabulary probability greater than a preset sampling threshold are selected. The words predicted in the previous time step are used as the known input for the current time step to continue the deduction. Multiple candidate text paths containing continuous task processing nodes are gradually generated by autoregressive sampling.

[0048] Next, the beam search algorithm is used to accumulate the single-step logarithmic probabilities of multiple candidate text paths at each time step to obtain the global generation confidence of each candidate text path. The candidate text paths are then sorted in descending order according to their generation confidence, and the top 5 candidate text paths are truncated and retained. These are then output as candidate task processing sequences to the subsequent evaluation module.

[0049] This example leverages the contextual modeling capabilities and autoregressive inference mechanism of a pre-trained large model to match the generated candidate task processing sequence with the prior map state of the target object. A beam search algorithm is used to retain candidate text paths with high confidence, controlling the number of candidate paths and providing candidate task processing sequences for subsequent constraint evaluation and comprehensive scoring.

[0050] As an example, the training steps of the pre-trained large model include: acquiring historical sample data and corresponding real task processing sequences to construct a supervised instruction fine-tuning dataset; converting the supervised instruction fine-tuning dataset into a standard mapping pair containing contextual cues and labeled task processing sequences; inputting the standard mapping pair into the initial large model to output a predicted task processing sequence; calculating the autoregressive cross-entropy loss value between the predicted task processing sequence and the labeled task processing sequence; and updating the network parameters of the initial large model in reverse using a gradient descent algorithm based on a preset learning rate and the autoregressive cross-entropy loss value until a first preset convergence condition is met, thereby obtaining the pre-trained large model.

[0051] It should be noted that, in this example, historical sample data refers to object status records, operation logs, processing records, or business text materials previously stored in the target business system. The real task processing sequence can be a set of multi-timestep task processing steps formed from historical sample data and verified manually or through system feedback. The supervised instruction fine-tuning dataset refers to a dataset used to train the language generation model, which is transformed by combining a large number of labeled contextual inputs with labeled task processing sequences. The initial large model refers to a language generation model that already possesses general language generation capabilities but has not yet undergone task adaptation training based on the supervised instruction fine-tuning dataset for the target business scenario; the model structure, word embedding layers, and candidate vocabulary of the initial large model can be consistent with the pre-trained large model, the difference being that its network parameters have not yet been updated to be adapted to the target business scenario.

[0052] The first preset convergence condition refers to the threshold standard set to determine whether the network training process has ended. For example, it could be that the absolute value of the difference between the loss function values ​​of two consecutive iterations is less than 0.001 for five consecutive training batches. The preset learning rate is a hyperparameter used in gradient descent calculations to control the step size of network weight parameter updates. It is typically set between 1e-5 and 5e-5 to prevent the step size from being too large, causing the model to fail to converge, or too small, causing the training cycle to be too long. The labeled task processing sequence refers to the set of standard task processing steps corresponding to historical sample data, which have been manually labeled, verified by system feedback, or confirmed by historical processing records. It is used as the target output label during supervised training. The predicted task processing sequence refers to the task processing text sequence generated step-by-step by the initial large model based on contextual cues. It is used to compare with the labeled task processing sequence and calculate the training loss value. The predicted task processing sequence is an intermediate output result during model training and is not the final target task processing sequence sent to the target business system. Autoregressive cross-entropy loss is a loss value calculated during the autoregressive generation process based on the difference between the vocabulary probability distribution and the corresponding labeled lexical units at each generation time step. It measures the deviation between the prediction task processing sequence and the labeling task processing sequence. A standard mapping pair refers to a training sample pair consisting of contextual cues and the corresponding labeled task processing sequence, where the contextual cues serve as the model input and the labeled task processing sequence serves as the model's expected output.

[0053] Understandably, the process begins by extracting historical sample data and corresponding real task processing sequences from the target business system's historical database. This raw business data is then cleaned and formatted to construct a supervised instruction fine-tuning dataset. This dataset is then processed using the model's instruction fine-tuning template format, involving text concatenation and variable substitution. Object state records from the historical sample data are extracted as input conditions to construct contextual cues, and processing nodes from the real task processing sequences are extracted as target outputs to construct labeled task processing sequences. This transforms the dataset into a standard mapping pair containing contextual cues and labeled task processing sequences.

[0054] The contextual cues contained in the standard mapping pairs are input into an initial large model that is either untuned or has only general pre-trained weights. This model utilizes its internal attention mechanism and decoding network to perform text inference and output a predicted task processing sequence. The probability distribution of the predicted task processing sequence along the vocabulary dimension is obtained, and the labeled lexical tags of the labeled task processing sequence at the corresponding time steps are extracted. These two are compared and substituted into the cross-entropy formula to calculate the autoregressive cross-entropy loss value between the predicted and labeled task processing sequences. The labeled lexical tags refer to the standard lexical indices corresponding to the labeled task processing sequence at each generation time step, used to calculate the cross-entropy loss with the probability distribution of the predicted task processing sequence along the vocabulary dimension.

[0055] Furthermore, the first preset convergence condition and preset learning rate are extracted. Based on the calculated autoregressive cross-entropy loss value, the partial derivatives of the loss function with respect to each network parameter in the initial large model are calculated using the AdamW gradient descent algorithm. Combined with the preset learning rate, the parameter update amount is calculated, and the network parameters of the initial large model are updated in reverse according to this update amount. The above forward deduction and reverse parameter update process is repeated continuously, monitoring the change of loss value in each training batch, until the decrease in loss value meets the first preset convergence condition. At this point, the network parameter update is stopped, the network weights are fixed, and the pre-trained large model is obtained.

[0056] This example utilizes real historical business data to construct instruction fine-tuning data, enabling an initial large model, which has not been adapted to a specific business scenario, to learn the task processing logic and sequence generation preferences in the target business scenario. By combining the supervision signal of the cross-entropy loss function with the gradient descent optimization mechanism, the underlying parameter weights of the language generation model are effectively adjusted, reducing the probability of the model outputting irrelevant or erroneous text when faced with complex heterogeneous data. This improves the pre-trained large model's ability to follow task instructions and the matching degree of generated content in task processing sequence generation scenarios.

[0057] As an example, the step of inputting each of the candidate task processing sequences into a pre-trained scoring model and extracting the temporal features of each of the candidate task processing sequences through the scoring model includes: embedding the fused prompt sequence as an object context, and inputting each of the candidate task processing sequences into the embedding layer of the pre-trained scoring model to transform them into candidate sequence embeddings; concatenating the object context embeddings and the candidate sequence embeddings along the temporal dimension to obtain a complete evaluation sequence containing object context and task processing content; using the complete evaluation sequence as the initial bottom-level vector of the scoring model; calculating the attention weights between the initial bottom-level vectors at different time steps through the multi-head self-attention layer in the scoring model; and performing feature aggregation and feedforward network mapping on the initial bottom-level vectors according to the attention weights to obtain the temporal features of each of the candidate task processing sequences.

[0058] It should be noted that object context embedding refers to a continuous vector representation formed by fusing cue sequences, used to characterize the state of the target object and task cue information. Candidate sequence embedding refers to a continuous vector representation obtained by mapping the candidate task processing sequence through the embedding layer of the scoring model. The complete evaluation sequence refers to the input representation of the scoring network formed by combining the object context embedding and the candidate sequence embedding along the temporal dimension. The initial bottom layer vector refers to the bottom layer dense vector representation corresponding to the complete evaluation sequence, used as the input feature of the multi-head self-attention layer. The multi-head self-attention layer refers to a neural network component containing multiple parallel computing branches, each branch capable of capturing the global dependencies of the input sequence from different representation subspaces.

[0059] Specifically, firstly, the fusion prompt sequence generated by the preceding process and the candidate task processing sequences output by the pre-trained large model are obtained, and the fusion prompt sequence is embedded as the object context; then, each candidate task processing sequence is split into multiple lexical units or task processing nodes, and transformed into candidate sequence embeddings through the embedding layer of the scoring model.

[0060] Next, the object context embedding and the candidate sequence embedding are concatenated in chronological order to obtain a complete evaluation sequence containing the object context and task processing content, and the complete evaluation sequence is used as the initial underlying vector of the scoring model.

[0061] Next, the initial bottom-level vectors are input into the multi-head self-attention layer of the scoring model. Multiple independent linear transformation matrices are used to map the initial bottom-level vectors into query tensors, key tensors, and value tensors, respectively. By calculating the dot product between the query tensor and the key tensor and applying Softmax normalization, the attention weights between the initial bottom-level vectors at different time steps are obtained. These attention weights numerically reflect the degree of correlation between the object's context state and subsequent task processing nodes within the same sequence. The query tensor, key tensor, and value tensor are intermediate tensors used in the multi-head self-attention layer to calculate attention weights and feature aggregation, respectively.

[0062] The calculated attention weights are used to perform a weighted summation of the corresponding value tensors, completing the initial feature aggregation of global information within the complete evaluation sequence. The aggregated tensor features are then input into a feedforward network containing a ReLU activation function to perform a nonlinear mapping transformation of the feature space. After processing through the internal residual connection structure, the hidden layer vectors corresponding to the global representation labels at the sequence head are extracted, or the features at all time steps in the sequence are subjected to average pooling, outputting the temporal features of each candidate task processing sequence after dimensionality reduction. Here, the residual connection structure refers to the connection structure that adds or fuses the network layer input and output. The global representation label refers to a specific sequence position label used to summarize the global semantic information of the entire input sequence.

[0063] This example pre-concatenates the object context with the candidate task processing sequence, enabling the evaluation network to perceive the object state premise corresponding to the current task processing sequence. By utilizing a mechanism combining multi-head self-attention layers and feedforward network mapping, the constraint correlations between task nodes at various time steps in the long sequence task processing can be mined, improving the global representation capability of sequence features and providing underlying feature support for subsequent quantification of sequence quality evaluation results and constraints.

[0064] As an example, the training steps of the scoring model include: extracting multiple comparison task processing sequences generated by the pre-trained large model based on the contextual cues for the same target object; sorting and labeling the multiple comparison task processing sequences based on preset sequence quality indicators and preset constraints to obtain sorting and labeling results, and constructing a partially ordered sequence dataset based on the sorting and labeling results; inputting the partially ordered sequence dataset into an initial scoring network, and calculating the predicted score value of each comparison task processing sequence through the regression output head in the initial scoring network; calculating the pairwise sorting loss value of the initial scoring network based on the sorting and labeling results and the predicted score value; updating the network weights of the initial scoring network according to the pairwise sorting loss value until a second preset convergence condition is met, thereby obtaining the trained scoring model.

[0065] It should be noted that contextual cues refer to the model input text or input vector generated based on historical sample data of the target object, object state records, task constraint information, or cue templates, used to guide the initial large model in generating a predicted task processing sequence. In this example, the comparative task processing sequence refers to the text path containing differentiated task processing steps generated by the pre-trained large model under different sampling parameters for the same preceding object context input. The predicted score is a one-dimensional scalar score output by the initial scoring network based on the deep global feature vector of the comparative task processing sequence, used to represent the relative quality evaluation result of the comparative task processing sequence under the current scoring network parameters. The deep global feature vector is the overall semantic feature representation obtained by the initial scoring network after encoding the comparative task processing sequence, used as input to the regression output head to calculate the predicted score. The ranking labeling result refers to the relative superiority / inferiority order labeling result formed for multiple comparative task processing sequences under the same contextual cues, used to indicate the ranking relationship between different comparative task processing sequences.

[0066] The pairwise ranking loss is a loss value calculated based on the ranking labeling results and predicted scores. It constrains the initial scoring network to output higher predicted scores for better task processing sequences and lower predicted scores for worse task processing sequences. The partially ordered sequence dataset can be a collection of samples containing relative superiority / inferiority relationships between sequences, typically organized in "winner-loser" pairs. The initial scoring network can be a pre-trained language model based on the RoBERTa architecture. The regression output head refers to the fully connected linear layer at the network's end that maps high-dimensional feature representations to one-dimensional scalar values. The second pre-set convergence condition is that the pairwise ranking accuracy on the validation set does not significantly improve over three consecutive training epochs.

[0067] Understandably, the process involves obtaining contextual clues for the same target object, calling a pre-trained large model, and generating multiple comparative task processing sequences based on these clues. Pre-defined sequence quality metrics and constraints are then obtained. The rationality and constraint fulfillment of the multiple comparative task processing sequences are evaluated against these metrics and constraints, and their relative ranking is assigned, resulting in a sorted and labeled dataset. Based on this sorted and labeled dataset, the different comparative task processing sequences generated for the same contextual clue are paired to distinguish between superior and inferior sequences, thus constructing a partially ordered sequence dataset.

[0068] Next, the processing sequences of each comparative task in the partially ordered sequence dataset are input into the initial scoring network, and the deep global feature vectors of the sequences are extracted using the attention encoder structure inside the network. The deep global feature vectors are then input into the regression output head to output the predicted scores of each comparative task processing sequence.

[0069] The true partial order relationship is extracted from the ranking annotation results. Combined with the predicted scores of each comparison task processing sequence, the pairwise ranking loss of the initial scoring network is calculated using the log-Sigmoid function. This loss reduces the probability that the predicted score of a worse sequence is higher than that of a better sequence. The gradient of the pairwise ranking loss with respect to the weights of each layer of the initial scoring network is calculated, and the network weights of the initial scoring network are updated using the AdamW gradient optimization algorithm. The above feature extraction and backpropagation process is repeated until the network performance meets the second preset convergence condition, at which point training stops, and the trained scoring model is obtained. The true partial order relationship refers to the relative superiority or inferiority relationship between the comparison task processing sequences recorded in the ranking annotation results, used to indicate the ranking order of different comparison task processing sequences under the same contextual cues.

[0070] This example establishes a feedback training mechanism based on relative merit ranking, overcoming the problem of inconsistent evaluation scales when directly scoring long text task processing sequences. By introducing sequence quality indicators and constraints that align with the target business scenario as guidance, the trained scoring model can measure the overall quality differences between different task processing sequences, providing a quantitative scoring basis for subsequently selecting target task processing sequences that meet the constraints.

[0071] As an example, the step of outputting a comprehensive score value for each candidate task processing sequence satisfying a preset sequence quality index and preset constraints through the scoring model based on the temporal features includes: inputting the temporal features into the sequence quality evaluation head of the scoring model to calculate the sequence fitting score of the temporal features under the preset sequence quality index; inputting the temporal features into the constraint penalty network in the scoring model to obtain the constraint violation probability of each candidate task processing sequence; if the constraint violation probability is greater than a preset constraint violation threshold, using a preset low score as the constraint satisfaction score; if the constraint violation probability is less than or equal to the preset constraint violation threshold, calculating the constraint satisfaction score based on the constraint violation probability, wherein the constraint satisfaction score is negatively correlated with the constraint violation probability; and performing a weighted summation of the sequence fitting score and the constraint satisfaction score based on a preset weighting parameter to obtain a comprehensive score value for each candidate task processing sequence satisfying the preset sequence quality index and the preset constraints.

[0072] It should be noted that in this example, the sequence quality assessment head can be a regression branch network built from a linear transformation layer at the end of the scoring model, used to map high-dimensional features to sequence quality scores. The constraint penalty network refers to a probabilistic output branch built from a multi-layer feedforward neural network, used to identify the probability that a candidate task processing sequence violates preset constraints. The sequence fitting score is a scalar score output by the sequence quality assessment head, characterizing the degree to which a candidate task processing sequence meets preset sequence quality indicators. The preset constraint violation threshold is a probability limit set to determine whether a candidate task processing sequence has a serious risk of constraint conflict; its value can be 0.9 to filter candidate task processing sequences that do not meet preset constraints. The preset low score can be a value used to reduce the ranking of risky sequences, for example, a value of 0. The preset weighting parameter can be a floating-point value used to balance the sequence quality indicator and constraints; for example, setting the sequence quality weight to 0.7 and the constraint satisfaction weight to 0.3. The constraint satisfaction score characterizes the degree to which a candidate task processing sequence meets preset constraints, and its value can be negatively correlated with the constraint violation probability.

[0073] Specifically, the temporal features are input into the sequence quality assessment head of the scoring model. The weight matrix inside the sequence quality assessment head is used to perform a linear mapping transformation on the temporal features, projecting the high-dimensional temporal features into a one-dimensional scalar numerical space, outputting a continuous scalar score, and using this scalar score as the sequence fitting score that characterizes the degree to which the current sequence achieves the preset sequence quality index.

[0074] Next, the temporal features are input into the constraint penalty network in the scoring model. The nonlinear classifier inside the constraint penalty network processes the features and uses the Sigmoid activation function to map the features into probability values ​​between 0 and 1. This probability value mathematically quantifies the likelihood that the temporal features will trigger the restrictive clauses in the preset constraints, and is used as the constraint violation probability of each candidate task processing sequence.

[0075] If the probability of constraint violation is greater than the preset constraint violation threshold, it indicates that the candidate task processing sequence has a serious risk of constraint conflict, and a preset low score is used as the constraint satisfaction score for the sequence. If the probability of constraint violation is less than or equal to the preset constraint violation threshold, it indicates that the probability of constraint violation has not exceeded the preset constraint violation threshold, and the constraint satisfaction score is calculated based on the constraint violation probability; for example, the difference between 1 and the constraint violation probability is used as the constraint satisfaction score.

[0076] Furthermore, the preset weighting parameters are obtained, and the corresponding weighting factors are multiplied by the sequence fitting score and constraint satisfaction score, respectively, and then summed to calculate the comprehensive score value of each candidate task processing sequence that satisfies the preset sequence quality index and preset constraint conditions. The formula for calculating the comprehensive score value is as follows: Where S represents the overall score; This represents the sequence fitting score; Indicates the constraint satisfaction score; and These represent the weight factors in the preset weighting parameters that correspond to sequence quality and constraints, respectively. .

[0077] In one example, constraint satisfaction score It can be calculated using the following formula: in, This represents the probability of constraint violation.

[0078] This example achieves dual-index quantification of candidate task processing sequences by constructing an independent sequence quality evaluation head and constraint penalty network in the scoring model. By introducing a segmented evaluation mechanism based on constraint violation probability and setting a preset low score when the constraint violation probability is high, the probability of selecting candidate task processing sequences that violate preset constraints can be reduced. When the constraint violation probability does not exceed the preset constraint violation threshold, a comprehensive score is calculated by combining the constraint satisfaction score and the sequence fitting score, which helps to improve the screening stability of the target task processing sequences.

[0079] As an example, the step of selecting a target task processing sequence from multiple candidate task processing sequences based on the comprehensive score value and sending the target task processing sequence to the target business system via a communication interface includes: sorting the multiple candidate task processing sequences in descending order according to the comprehensive score value and extracting the first sequence; if the comprehensive score value of the first sequence is greater than or equal to a preset backoff threshold, using the first sequence as the target task processing sequence; if the comprehensive score value of the first sequence is less than the preset backoff threshold, extracting object type identifiers from the multi-source heterogeneous data; matching the object type identifiers with rule indexes in a preset rule base to retrieve associated preset rule sequence paths, which are then used as the target task processing sequence; and converting the target task processing sequence into a visual timeline chart format and sending it to the target business system for display via a communication interface.

[0080] It should be noted that in this example, the preset backoff threshold can be a benchmark value used to determine whether the candidate task processing sequence output by the pre-trained large model meets the preset output conditions. Its value can be 0.6, which can be determined based on historical test data statistics to balance the sequence adoption rate and constraint risk. Preset output conditions can include a comprehensive score not lower than the preset backoff threshold, no serious constraint conflicts triggered, and the output format meeting the target business system interface requirements. Object type identifiers refer to the target object category, task type, fault type, or business event type recorded in multi-source heterogeneous data. The preset rule base can be an internally pre-built structured database storing fixed processing rules corresponding to various object types. The rule index refers to a keyword mapping table used to quickly locate the corresponding preset rule sequence path in the preset rule base. The preset rule sequence path can be a fixed and deterministic sequence of task processing steps that does not depend on the pre-trained large model. The visualization timeline chart format refers to a data presentation structure that displays task processing steps in chronological order using graphical nodes and connections.

[0081] Understandably, the process begins by extracting the comprehensive score value corresponding to each candidate task processing sequence. A sorting algorithm is then used to rank the candidate sequences in descending order of their comprehensive scores, and the sequence ranked first is selected as the first-order sequence. Next, it is determined whether the comprehensive score value of the first-order sequence is greater than or equal to a preset backoff threshold of 0.6. If the comprehensive score value of the first-order sequence is greater than or equal to the preset backoff threshold, it indicates that the current first-order sequence meets the preset output condition, and the first-order sequence is directly used as the target task processing sequence.

[0082] If the overall score of the first sequence is less than a preset backoff threshold, it indicates that none of the candidate task processing sequences generated by the pre-trained large model meet the preset output conditions. In this case, a backoff mechanism is triggered, using text matching rules to extract the recorded object type identifier from the multi-source heterogeneous data of the target object. The backoff mechanism means that when the overall score of the first sequence is lower than the preset backoff threshold, the candidate task processing sequence generated by the pre-trained large model is not used; instead, a preset rule sequence path is retrieved from the preset rule base as the target task processing sequence.

[0083] Next, the object type identifier is compared and matched with the rule index inside the preset rule base by keywords. The preset rule sequence path that completely corresponds to the object type identifier is queried and retrieved. The retrieved preset rule sequence path replaces the pre-trained large model generation sequence as the target task processing sequence.

[0084] The process extracts each task processing node and its corresponding time node from the target task processing sequence. A graphics rendering component maps these nodes to their corresponding positions on a time axis, transforming the sequence into a visual timeline chart. This chart is then sent to the front-end terminal device of the target business system for display via a communication interface. The graphics rendering component can be a program module on the front end of the target business system used to generate the timeline chart. The communication interface can be the interface between the electronic device and the target business system for transmitting the target task processing sequence or visual chart data.

[0085] This example effectively intercepts low-quality or potentially conflicting candidate task processing sequences by introducing a dual criterion of ranking based on comprehensive scores and a preset backoff threshold. When the model output does not meet preset output conditions, a preset rule sequence path is introduced as a backoff scheme, improving the robustness and output stability of the sequence generation scoring system. Combined with the data presentation method of a visual timeline, it reduces the parsing cost for the target business system to handle complex long-sequence task processing schemes.

[0086] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the sequence generation and scoring method based on a large model according to this application. Step S10 of the sequence generation and scoring method based on a large model includes steps S11 to S16: Step S11: Separate structured attribute data and unstructured text data from the multi-source heterogeneous data of the target object; Step S12: Identify target entity words in the unstructured text data using a preset entity extraction model, and convert the target entity words into semantic feature vectors using a preset text encoder; Step S13: Calculate the cosine similarity between the semantic feature vector and each graph node in the preset knowledge graph, and determine the graph nodes with a cosine similarity greater than the preset similarity threshold as target mapping nodes; Step S14: Discretize the structured attribute data into attribute features and attach and fuse them to the corresponding target mapping node; Step S15: Extract the multi-hop relationship edges and adjacent nodes of the target mapping node from the preset knowledge graph; Step S16: Based on the multi-hop relationship edge, perform feature aggregation on the target mapping node and the adjacent node that fuse the attribute features through a graph neural network to obtain the object subgraph features.

[0087] It should be noted that structured attribute data refers to numerical data with fixed fields, fixed formats, or clear dimensions in the target business system, such as device status parameters, resource usage parameters, interface call statistics, or business indicator data. Unstructured text data can be natural language records such as log text, work order descriptions, event descriptions, or manual notes recorded in the target business system. The preset entity extraction model can be a sequence labeling network built on the BiLSTM-CRF architecture. The preset text encoder can be a BERT language model pre-trained on a large-scale general corpus or target business corpus. The preset similarity threshold is a floating-point lower limit set to determine whether feature vectors and graph nodes belong to the same semantic concept. Its value is usually between 0.8 and 0.95. In this example, it is specifically set to 0.85 to balance the recall and precision of entity mapping. Attribute features refer to discrete category labels obtained by dividing continuous numerical attribute data into different intervals according to preset business thresholds. Mounting and fusion refers to using attribute features as additional attributes, node labels, or associated features of the target mapping nodes, participating together with the target mapping nodes in the subsequent feature aggregation of the graph neural network.

[0088] A multi-hop relation edge refers to a graph edge obtained by extending outward from the target mapping node along the knowledge graph connection relationship within a range of two or more hops. The graph neural network can be a graph convolutional network, a graph attention network, or other neural networks used for feature aggregation of graph structured data. Target entity words refer to keywords or phrases identified from unstructured text data that are related to the state, task type, business event, or processing action of the target object. Semantic feature vectors refer to the continuous vector representation obtained by encoding target entity words through a pre-defined text encoder. Target mapping nodes refer to graph nodes in the pre-defined knowledge graph that successfully match target entity words or structured attribute data. Adjacent nodes refer to graph nodes in the pre-defined knowledge graph that are directly connected to the target mapping node through a one-hop relation edge.

[0089] Specifically, firstly, the multi-source heterogeneous data of the target object is classified and extracted according to the data storage format. Data with explicit fields and numerical formats, such as device temperature, resource utilization, and interface error rate, are separated into structured attribute data. Natural language paragraphs containing log records, work order descriptions, or event descriptions are separated into unstructured text data. The unstructured text data is input into a preset entity extraction model, which uses a conditional random field layer within the model to perform label decoding on the text sequence, identifying and extracting target entity words such as object names, exception types, processing actions, or business events contained in the text. The extracted target entity words are then input into a preset text encoder, which outputs semantic feature vectors with fixed dimensions.

[0090] Extract the initial node vectors corresponding to each graph node in the preset knowledge graph, and calculate the cosine similarity between the semantic feature vector and each initial node vector using the vector dot product. The initial node vector is a vectorized representation used to characterize the semantic, attribute, or relational context information of the graph nodes in the preset knowledge graph; it can be encoded by node name, node attribute, node type, or node connection relationship. Extract a preset similarity threshold of 0.85, and filter graph nodes with a cosine similarity greater than 0.85 and determine them as target mapping nodes. Obtain the separated structured attribute data, and discretize the continuous values ​​into attribute features such as high, normal, low, or abnormal, referring to a preset business threshold range. The discretized attribute features are then used as supplementary node information and fused into the corresponding target mapping nodes.

[0091] Next, within the pre-defined knowledge graph network structure, starting from the target mapping node, a search is performed outwards to extract relation edges and associated neighboring nodes within a two- or three-hop range. Based on the extracted multi-hop relation edges, a graph convolutional network is used to perform a weighted summation of neighborhood information on the hidden layer representations of the target mapping node and its neighboring nodes carrying attribute features, thereby completing the feature aggregation of the local network topology and outputting the dimensionality-reduced object subgraph features.

[0092] This embodiment realizes the transformation of multimodal data of target objects into structured network features. By separating structured attribute data and unstructured text data, it uses cosine similarity matching to complete the semantic alignment of graph nodes, solving the matching omission problem caused by the non-standard entity representation in unstructured text. The structured attribute data is discretized and mounted into the graph network, preserving the state categories corresponding to numerical attributes. The aggregation mechanism of the graph convolutional network is used to capture the association between multi-hop nodes, improving the feature representation effect of the target object's state and constraints, and providing pre-graph feature support for the subsequent pre-trained large model to generate candidate task processing sequences.

[0093] In one optional medical application scenario, the target object can be the target patient, and the multi-source heterogeneous data can include structured clinical indicators and unstructured medical record text that have been authorized for collection or storage. The candidate task processing sequence can be a candidate clinical decision reference sequence, and the target business system can be a hospital information system, an electronic medical record system, or a clinical information management system. The preset knowledge graph can be a knowledge graph built for a specific disease. Graph nodes can include symptom nodes, examination item nodes, laboratory indicator nodes, drug category nodes, contraindication rule nodes, follow-up item nodes, and guideline item nodes. Graph edges can be used to represent the association, dependency, or constraint relationships between nodes.

[0094] In this medical application scenario, historical sample data can include historical electronic medical records for specific diseases, laboratory test records, follow-up records, and anonymized clinical text data. Real-world task processing sequences can include corresponding historical clinical auxiliary reference paths, follow-up item sequences, or manually confirmed prompt sequence sequences. This type of data can be used to construct supervised instruction fine-tuning datasets or partially ordered sequence datasets to train pre-trained large models or scoring models.

[0095] In this medical application scenario, the target task processing sequence output to the medical information system for display is essentially only intended as auxiliary reference data and prompts for clinicians when formulating treatment plans. This system and its corresponding methods do not possess final clinical intervention decision-making authority. The comprehensive score is only used for ranking and filtering candidate task processing sequences and does not represent actual clinical efficacy, treatment benefits, or final medical conclusions. The target task processing sequences, clinical pathway references, or preset rule sequences provided in this example do not constitute a final diagnostic conclusion for a specific disease, do not directly trigger the issuance of legally binding prescriptions, and do not directly implement physical medical interventions. The final disease status assessment conclusion, treatment plan, or medical intervention plan should still be independently made by a qualified physician based on the patient's actual clinical condition.

[0096] For ease of understanding, the following example of type 2 diabetes follow-up management illustrates the sequence generation and scoring method based on a large model. This example is only used to explain one manifestation of the target object, multi-source heterogeneous data, candidate task processing sequences, and target business system in a medical scenario, and does not limit the application area of ​​this application.

[0097] For the target patient, the electronic device acquires multi-source heterogeneous data from the target business system. Structured attribute data may include glycated hemoglobin (HbA1c) values, fasting blood glucose values, blood pressure values, the date of the most recent fundus examination, the date of the most recent renal function test, and the current medication category; unstructured text data may include outpatient complaint text, medical history summary, examination report description, or follow-up remarks. For example, structured attribute data may include HbA1c of 8.5%, fasting blood glucose (FBG) of 9.2 mmol / L, and systolic blood pressure of 145 mmHg; unstructured text data may include an outpatient complaint text record stating, "The patient reports recent blurred vision, accompanied by symmetrical numbness in both lower extremities."

[0098] In step S10, the electronic device identifies target entity words in unstructured text data, such as "blurred vision" and "lower limb numbness," using a preset entity extraction model, and converts these target entity words into semantic feature vectors using a preset text encoder. The electronic device calculates the cosine similarity between the semantic feature vectors and each node in a preset knowledge graph, identifying nodes with similarity greater than a preset similarity threshold as target mapping nodes. Simultaneously, structured attribute data is discretized into attribute features according to a preset threshold and integrated into the corresponding target mapping nodes. For example, glycated hemoglobin values ​​are discretized into a blood glucose control status attribute, and the time of the most recent fundus examination is discretized into an examination interval status attribute. Afterward, the electronic device uses a graph neural network to aggregate features from target mapping nodes, adjacent nodes, and multi-hop relationship edges to obtain object subgraph features.

[0099] In step S20, the electronic device inputs the object subgraph features into a preset feature projection layer, and obtains a soft cue vector with the same dimension as the input space of the pre-trained large model through linear mapping and normalization. Simultaneously, it converts the template text fragments in the preset task cue template into text embedding vectors. The preset task cue template may include output format requirements, follow-up item generation requirements, constraint verification requirements, and cue rules prohibiting the direct generation of final diagnostic conclusions or prescription execution instructions. The electronic device concatenates the soft cue vector and the text embedding vector along the temporal dimension to obtain a fused cue sequence.

[0100] In step S30, the electronic device inputs the fused prompt sequence into a pre-trained large model for autoregressive decoding, generating multiple candidate task processing sequences. In this medical application scenario, the candidate task processing sequences can be represented as candidate clinical decision reference sequences. For example, the candidate task processing sequences may include reference items arranged in chronological order, such as "prompt to review current blood glucose control status -> prompt to assess the need for fundus examination -> prompt to assess the need for neurological symptom-related examinations -> prompt to set follow-up time." The aforementioned candidate task processing sequences serve only as information prompt sequences in the target business system and do not directly constitute disease diagnosis conclusions, treatment plans, prescription content, or medical operation instructions.

[0101] In step S50, the electronic device inputs each candidate task processing sequence into the scoring model, and extracts the temporal features of each candidate task processing sequence through the scoring model. The sequence quality assessment head in the scoring model calculates the sequence fitting score based on the temporal features, and the constraint penalty network calculates the constraint violation probability based on the temporal features. Preset sequence quality indicators may include the completeness of the event, the rationality of the time sequence, the degree of matching with the state of the target object, and the consistency with the labeling results of historical similar samples. Preset constraints may include rules for drug contraindication prompts, rules for conflicting examination items, rules for output format, rules for permission verification, and output restriction rules that prohibit making final medical judgments in place of licensed physicians.

[0102] For example, the constraint penalty network in the scoring model performs constraint verification on the candidate task processing sequence. If the sequence fails to trigger preset constraints, the constraint violation probability is output as 0.02, which is lower than the preset constraint violation threshold of 0.9. The electronic device calculates the constraint satisfaction score based on the constraint violation probability; for example, the constraint satisfaction score is 1 - 0.02 = 0.98. Simultaneously, the sequence quality assessment head calculates the sequence fit score of the candidate task processing sequence based on temporal characteristics, considering sequence quality indicators such as event completeness, temporal order rationality, and matching degree with the target patient's state. If the sequence fit score is 0.84, the sequence quality weight is 0.7, and the constraint satisfaction weight is 0.3, then the comprehensive score is 0.7 × 0.84 + 0.3 × 0.98 = 0.882, which the electronic device can record as 0.88.

[0103] In step S60, the electronic device sorts multiple candidate task processing sequences in descending order according to their comprehensive score and extracts the first sequence. If the comprehensive score of the first sequence is greater than or equal to a preset backoff threshold of 0.6, the first sequence is determined as the target task processing sequence. If the comprehensive score of the first sequence is less than the preset backoff threshold, the electronic device extracts object type identifiers from multi-source heterogeneous data and matches them with rule indexes in a preset rule base to retrieve the associated preset rule sequence path as the target task processing sequence. For example, the object type identifier could correspond to "type 2 diabetes follow-up management," and the preset rule sequence path could be a fixed follow-up item template stored in the preset rule base.

[0104] Electronic devices send the target task processing sequence to the target business system via a communication interface and can convert the sequence into a visual timeline chart for display. This timeline chart format can include reference nodes, suggested review time nodes, constraint prompt nodes, and a manual confirmation entry point. For example, the timeline can display reference nodes such as "Day 1: Review blood glucose control status," "Day 3: Assess fundus examination needs," and "Day 30: Review relevant indicators." After receiving the target task processing sequence, the target business system can display it to users with appropriate permissions for review.

[0105] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the sequence generation scoring method based on the large model of this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0106] This application also provides a sequence generation and scoring device based on a large model, please refer to... Figure 3 The sequence generation and scoring device based on the large model includes: Mapping module 10 is used to acquire multi-source heterogeneous data of target objects collected or stored by the target business system, perform node mapping and feature fusion on the multi-source heterogeneous data through a preset knowledge graph, and perform node feature aggregation through a graph neural network to obtain object subgraph features; The splicing module 20 is used to convert the object subgraph features into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and splice them with the text embedding vectors corresponding to the preset task cue template to obtain a fused cue sequence. Decoding module 30 is used to input the fusion prompt sequence into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences; The feature extraction module 40 is used to input each of the candidate task processing sequences into a pre-trained scoring model, and extract the temporal features of each of the candidate task processing sequences through the scoring model. The model output module 50 is used to output a comprehensive score value for each candidate task processing sequence that meets the preset sequence quality index and preset constraints based on the temporal features and through the scoring model. The filtering and sending module 60 is used to filter out the target task processing sequence from multiple candidate task processing sequences based on the comprehensive score value, and send the target task processing sequence to the target business system through a communication interface.

[0107] The large-model-based sequence generation and scoring device provided in this application, employing the large-model-based sequence generation and scoring method described in the above embodiments, can solve the technical problem of how to generate target task processing sequences that balance sequence quality indicators and constraints based on multi-source heterogeneous data. Compared with the prior art, the beneficial effects of the large-model-based sequence generation and scoring device provided in this application are the same as those of the large-model-based sequence generation and scoring method provided in the above embodiments, and other technical features in the large-model-based sequence generation and scoring device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0108] This application provides a sequence generation and scoring device based on a large model. The sequence generation and scoring device based on a large model includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the sequence generation and scoring method based on the large model in the first embodiment described above.

[0109] The following is for reference. Figure 4 This document illustrates a schematic diagram of a large-model-based sequence generation and scoring device suitable for implementing embodiments of this application. The large-model-based sequence generation and scoring device in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Android Devices), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The sequence generation and scoring device based on a large model shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0110] like Figure 4As shown, the large-model-based sequence generation and scoring device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM 1002 (Read Only Memory) or a program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the large-model-based sequence generation and scoring device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touch screens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the large-model-based sequence generation and scoring device to communicate wirelessly or wiredly with other devices to exchange data. While the figures show large-model-based sequence generation and scoring devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0111] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0112] The large-model-based sequence generation and scoring device provided in this application, employing the large-model-based sequence generation and scoring method described in the above embodiments, can solve the technical problem of how to generate target task processing sequences that balance sequence quality indicators and constraints based on multi-source heterogeneous data. Compared with the prior art, the beneficial effects of the large-model-based sequence generation and scoring device provided in this application are the same as those of the large-model-based sequence generation and scoring method provided in the above embodiments, and other technical features in this large-model-based sequence generation and scoring device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0113] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0114] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the sequence generation and scoring method based on a large model in the above embodiments.

[0115] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0116] The aforementioned computer-readable storage medium may be included in a large-model-based sequence generation and scoring device; or it may exist independently and not assembled into a large-model-based sequence generation and scoring device.

[0117] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0118] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0119] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0120] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described large-model-based sequence generation and scoring method. This solves the technical problem of how to generate target task processing sequences that balance sequence quality indicators and constraints based on multi-source heterogeneous data. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the large-model-based sequence generation and scoring method provided in the above embodiments, and will not be repeated here.

[0121] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the sequence generation and scoring method based on a large model as described above.

[0122] The computer program product provided in this application can solve the technical problem of how to generate target task processing sequences that take into account both sequence quality indicators and constraints based on multi-source heterogeneous data. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the sequence generation and scoring method based on large models provided in the above embodiments, and will not be repeated here.

[0123] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A large model-based sequence generation scoring method, characterized in that, The method is performed by an electronic device, and the method includes: Acquire multi-source heterogeneous data of target objects collected or stored by the target business system, perform node mapping and feature fusion on the multi-source heterogeneous data through a preset knowledge graph, and aggregate node features through a graph neural network to obtain object subgraph features; The object subgraph features are transformed into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and then concatenated with the text embedding vectors corresponding to the preset task cue template to obtain a fused cue sequence. The fusion prompt sequence is input into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences; Each of the candidate task processing sequences is input into a pre-trained scoring model, and the temporal features of each candidate task processing sequence are extracted through the scoring model. Based on the aforementioned temporal characteristics, the scoring model outputs a comprehensive score value for each candidate task processing sequence that satisfies the preset sequence quality index and preset constraints. The target task processing sequence is selected from multiple candidate task processing sequences based on the comprehensive score, and then sent to the target business system through a communication interface.

2. The method of claim 1, wherein, The steps of acquiring multi-source heterogeneous data of the target object collected or stored by the target business system, performing node mapping and feature fusion on the multi-source heterogeneous data through a preset knowledge graph, and aggregating node features through a graph neural network to obtain object subgraph features include: Separate structured attribute data and unstructured text data from multi-source heterogeneous data of the target object; The target entity words in the unstructured text data are identified by a preset entity extraction model, and the target entity words are converted into semantic feature vectors by a preset text encoder. Calculate the cosine similarity between the semantic feature vector and each graph node in the preset knowledge graph, and determine the graph nodes with a cosine similarity greater than a preset similarity threshold as target mapping nodes; The structured attribute data is discretized into attribute features and then mounted and fused into the corresponding target mapping node; Extract the multi-hop relationship edges and adjacent nodes of the target mapping node from the preset knowledge graph; Based on the multi-hop relationship edges, feature aggregation is performed on the target mapping node and the adjacent nodes that fuse the attribute features through a graph neural network to obtain object subgraph features.

3. The method as described in claim 1, characterized in that, The step of outputting a comprehensive score value for each candidate task processing sequence that satisfies the preset sequence quality index and preset constraints based on the temporal characteristics and through the scoring model includes: The time series features are input into the sequence quality assessment head of the scoring model to calculate the sequence fitting score of the time series features under the preset sequence quality index; The temporal features are input into the constraint penalty network in the scoring model to obtain the constraint violation probability of each candidate task processing sequence; If the probability of constraint violation is greater than a preset constraint violation threshold, a preset low score will be used as the constraint satisfaction score. When the constraint violation probability is less than or equal to the preset constraint violation threshold, a constraint satisfaction score is calculated based on the constraint violation probability, wherein the constraint satisfaction score is negatively correlated with the constraint violation probability; The sequence fitting score and the constraint satisfaction score are weighted and summed based on preset weighting parameters to obtain a comprehensive score value for each candidate task processing sequence that satisfies the preset sequence quality index and the preset constraint conditions.

4. The method as described in claim 1, characterized in that, The step of selecting the target task processing sequence from multiple candidate task processing sequences based on the comprehensive score and sending the target task processing sequence to the target business system through a communication interface includes: The candidate task processing sequences are sorted in descending order according to the comprehensive score, and the first sequence is extracted. If the overall score of the first sequence is greater than or equal to a preset backoff threshold, the first sequence will be used as the target task processing sequence. If the overall score of the first sequence is less than the preset backoff threshold, the object type identifier is extracted from the multi-source heterogeneous data; The object type identifier is matched with the rule index in the preset rule base to retrieve the associated preset rule sequence path, which is used as the target task processing sequence. The target task processing sequence is converted into a visual timeline chart format and sent to the target business system for display via a communication interface.

5. The method as described in claim 1, characterized in that, The step of transforming the object subgraph features into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and concatenating them with the text embedding vectors corresponding to the preset task cue template to obtain the fused cue sequence includes: The object subgraph features are linearly mapped by a preset feature projection layer to obtain an initial continuous vector with the same input space dimension as the pre-trained large model. The initial continuous vector is normalized to obtain a soft cue vector aligned with the feature scale. Extract template text fragments from preset task prompt templates, and convert the template text fragments into text embedding vectors through a preset word embedding layer; The soft cue vector and the text embedding vector are concatenated along the temporal dimension to obtain the fused cue sequence.

6. The method as described in claim 1, characterized in that, The step of inputting each of the candidate task processing sequences into a pre-trained scoring model and extracting the temporal features of each of the candidate task processing sequences through the scoring model includes: The fusion prompt sequence is embedded as an object context, and each candidate task processing sequence is input into the embedding layer of a pre-trained scoring model to be transformed into a candidate sequence embedding. The object context embedding and the candidate sequence embedding are concatenated along the temporal dimension to obtain a complete evaluation sequence containing the object context and task processing content; The complete evaluation sequence is used as the initial underlying vector of the scoring model; The attention weights between the initial bottom-level vectors at different time steps are calculated through the multi-head self-attention layer in the scoring model. Based on the attention weights, feature aggregation and feedforward network mapping are performed on the initial underlying vectors to obtain the temporal features of each candidate task processing sequence.

7. The method according to any one of claims 1 to 6, characterized in that, The step of inputting the fused prompt sequence into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences includes: The fusion cue sequence is input into the feature extraction block of the pre-trained large model to extract the contextual hidden state; The contextual hidden state is mapped step-by-step by time using the language decoding head in the pre-trained large model, and the vocabulary probability distribution at each time step is output. Based on the probability distribution of the vocabulary at each time step, multiple candidate text paths are generated by autoregressive sampling from the candidate vocabulary of the pre-trained large model. The generation confidence of each candidate text path is calculated using a beam search algorithm; The candidate text paths are sorted in descending order according to the generated confidence level, and the first preset number of candidate text paths are taken as the candidate task processing sequence.

8. A sequence generation and scoring device based on a large model, characterized in that, The apparatus employs the sequence generation and scoring method based on a large model as described in any one of claims 1 to 7, and the apparatus comprises: The mapping module is used to acquire multi-source heterogeneous data of target objects collected or stored by the target business system, perform node mapping and feature fusion on the multi-source heterogeneous data through a preset knowledge graph, and aggregate node features through a graph neural network to obtain object subgraph features. The concatenation module is used to transform the features of the object subgraph into soft cue vectors with the same input space dimension as the pre-trained large model through a preset feature projection layer, and concatenate them with the text embedding vectors corresponding to the preset task cue template to obtain a fused cue sequence. The decoding module is used to input the fused prompt sequence into the pre-trained large model for autoregressive decoding to generate multiple candidate task processing sequences; The feature extraction module is used to input each of the candidate task processing sequences into a pre-trained scoring model, and extract the temporal features of each of the candidate task processing sequences through the scoring model. The model output module is used to output a comprehensive score value for each candidate task processing sequence that meets the preset sequence quality index and preset constraints based on the temporal features and through the scoring model. The filtering and sending module is used to filter out the target task processing sequence from multiple candidate task processing sequences based on the comprehensive score value, and send the target task processing sequence to the target business system through a communication interface.

9. A sequence generation and scoring device based on a large model, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the sequence generation and scoring method based on a large model as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the sequence generation and scoring method based on a large model as described in any one of claims 1 to 7.

Images