System and method for uniformly representing event description and numerical characteristics based on large model

Abstract

The invention provides an event description and numerical feature unified representation system and method based on a large model, belongs to the technical field of artificial intelligence, natural language processing and large models, and is used for solving the problem that cross-modal alignment and unified understanding are difficult to realize due to splitting of event semantics and numerical feature representation in the traditional technology. The method comprises the following steps: step A, carrying out text semantic embedding on natural language description of a risk event to generate semantic vector representation; b, semantic embedding is carried out on the numerical features of the risk events, and numerical vector representation is generated; c, modeling based on risk event Schema, and aligning and fusing semantic vectors and numerical vectors according to structured slots; and D, realizing consistent representation of semantic vectors and numerical vectors in a unified vector space through joint training and comparative learning, and supporting risk event understanding, prediction and decision-making tasks in the fields of intelligent traffic, disaster early warning and the like.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, natural language processing and large model technology, specifically involving a system and method for unified representation of event description and numerical features based on large models. Background Technology

[0002] With the widespread application of Large Language Models (LLMs) in tasks such as natural language understanding, event analysis, and prediction, a key challenge in handling risk events is enabling models to simultaneously understand the semantic description and related numerical features (such as time, intensity, and probability). Describing risk events primarily involves describing dynamic processes, encompassing rich textual information (e.g., "Severe traffic congestion occurred at a certain intersection on December 22, 2025") as well as specific numerical attributes (e.g., congestion start time, duration, impact area, and traffic flow change rate). Especially in scenarios such as smart cities, emergency management, and natural disaster analysis, the ability to achieve a unified understanding of event semantics and numerical values ​​directly impacts the accuracy of system analysis, the timeliness of response, and the interpretability of decisions.

[0003] Against this backdrop, leveraging large models to achieve multi-dimensional representation and reasoning of risk events has become a research hotspot. Some research institutions and teams have made initial progress in this direction. For example, the iTransformer model proposed by the research team at the School of Software, Tsinghua University, reconstructs the attention mechanism of Transformer and effectively captures the correlation between multiple variables by inverting the time series dimension, achieving good performance in numerical sequence prediction tasks. Zhejiang University and Ant Group, in their EventRAG research, proposed an enhanced generative framework based on event knowledge graphs. By transforming unstructured documents into event graph nodes containing time dependencies and logical relationships, and utilizing agents for iterative retrieval and reasoning, they significantly improved the logical consistency and long-range reasoning capabilities of large models in complex narrative scenarios. However, existing methods still have significant limitations in achieving a unified representation of event descriptions and numerical features. Common technical paths and their shortcomings are as follows: (1) Text semantic embedding method; This method directly encodes the event description text using a pre-trained language model to generate semantic vectors. Numerical features (such as "lasts 30 minutes") are treated as ordinary text tokens.

[0004] Disadvantages: It is impossible to specifically model the mathematical characteristics and semantics of numerical values, resulting in a weak correlation between "30 minutes" and "severe congestion" in the vector space, making it difficult to support downstream tasks such as numerical comparison and quantitative reasoning.

[0005] (2) Post-fusion multimodal method; The text is processed using a language model and the numerical values ​​are processed using a dedicated encoder. After obtaining two types of vectors, they are fused by vector concatenation, weighted addition, or a simple attention mechanism.

[0006] Disadvantages: Text and numerical vectors are usually generated independently in the early stages of training, lacking explicit constraints for cross-modal alignment, resulting in "semantic-numerical disconnect". The text vector describing "intensified congestion" and the numerical vector describing "speed decrease rate of 0.8" are far apart in space, affecting the consistency of the overall event representation.

[0007] (3) Rule- or template-driven structured methods; A schema is predefined for the event, including fixed slots such as time, location, and intensity. Text and numerical values ​​are then filled into the corresponding slots using information extraction technology to form a structured record.

[0008] Disadvantages: It relies on manually designed schemas, has poor generalization ability, only achieves alignment at the symbol level, and does not achieve deep integration of semantics and numerical values ​​at the vector representation level, making it difficult to support complex reasoning based on representations.

[0009] While existing methods have advanced event understanding research at different levels, they generally suffer from core problems such as the separation of semantic and numerical representations, weak alignment mechanisms, and the lack of a unified representation framework. This limits the performance of large models when dealing with complex tasks that require simultaneously understanding "what happened" and "how much, when, where, and how" an event evolves, hindering true cross-modal event understanding and reasoning and restricting their application in complex scenarios such as intelligent decision-making and real-time situational awareness. Therefore, researching new methods that can achieve deep integration and unified representation of event language descriptions and numerical features has significant theoretical and practical value. Summary of the Invention

[0010] To address the problems mentioned in the background, this invention provides a system and method for unified representation of event description and numerical features based on a large model, thereby solving the problems of semantic and numerical representation being separated, weak alignment mechanisms, and lack of a unified representation framework in existing technologies.

[0011] To achieve the above objectives, the present invention provides the following technical solution: The system based on large model event description and unified representation of numerical features includes a text semantic embedding module with a pre-trained large language model, a numerical feature embedding module with a numerical encoder, a schema modeling and fusion module, and a joint training and alignment module. The text semantic embedding module is used to semantically embed the natural language description of risk events, generate semantic vectors, and output them to the schema modeling and fusion module; The numerical feature embedding module is used to semantically embed the numerical features of risk events, generate numerical vectors, and output them to the schema modeling and fusion module. The Schema modeling and fusion module is used to receive semantic vectors and numerical vectors, and to perform structured alignment and fusion of the received data based on event schema modeling to generate unified event representation vectors. The joint training and alignment module establishes bidirectional data interaction with the text semantic embedding module, the numerical feature embedding module, and the schema modeling and fusion module, respectively. Through joint training and alignment, semantic vectors and numerical vectors achieve semantically consistent representation in a unified vector space.

[0012] The method for unified representation of event description and numerical features based on large models includes the following steps: Step A: Semantically embed the natural language description of the risk event to generate a semantic vector; Step B: Semantically embed the numerical features of the risk event to generate a numerical vector; Step C: Based on event schema modeling, the semantic vector and numerical vector are structurally aligned and fused to generate a unified event representation vector; Step D: Based on the event representation vector, through joint training and alignment, the semantic vector and the numerical vector achieve semantically consistent representation in a unified vector space.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This application does not rely on fixed event templates or complex multi-pipeline splicing. Instead, it uses the core mechanisms of "semantic-numerical joint embedding" and "structural binding-joint alignment" to enable large models to learn from unstructured text and structured numerical data together, autonomously constructing a fusion representation of risk events in a unified vector space. This enables a collaborative understanding and deep cognition of the essence of the event (what happened) and its quantitative attributes (how severe and when it happened).

[0014] This application achieves deep integration of semantic and numerical features of risk events, enhancing the ability of large models to understand risk events; it provides a structured representation framework through event schema, enhancing the semantic consistency of event representation; it supports risk event modeling and is applicable to event analysis and prediction in scenarios such as traffic and disasters; and it reduces training costs and improves model generalization ability through parameter fine-tuning and comparative learning. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for unifying event description and numerical feature representation based on a large model. Figure 2A detailed diagram illustrating the structured alignment and fusion process of semantic-numerical features under an event schema; Figure 3 A schematic diagram illustrating the alignment principle of semantic vectors and numerical vectors with the vector space during joint training, as provided in this invention. Figure 4 and Figure 5 This is a diagram illustrating the learning loss for comparison. Detailed Implementation

[0016] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0017] like Figure 1 As shown, the method for unified representation of event description and numerical features based on a large model includes the following steps: Step A: Semantically embed the natural language description of the risk event to generate a semantic vector representation; Step B: Semantically embed the numerical features of the risk event to generate a numerical vector representation; Step C: Based on event schema modeling, the semantic vector and numerical vector are structurally aligned and fused to generate a unified event representation vector; Step D: Based on the event representation vector, through joint training and alignment, the semantic vector and the numerical vector achieve semantically consistent representation in a unified vector space.

[0018] Step A specifically includes: Step A1: Encode the natural language description text of the event using a pre-trained language model based on the Transformer architecture, capture its contextual dependencies, and generate basic semantic vectors; Step A2: Identify and enhance the core elements in the event description by adding... <trigger> 、 <arg-time>Special markers are used to enhance vectors for event trigger words and arguments such as time and location. Step A3: Fine-tune the pre-trained language model using event corpora from specific domains to improve the relevance and accuracy of semantic embeddings in target domains such as transportation and disasters.

[0019] Step B specifically includes: Step B1: For time-series features such as timestamps and durations, use sine / cosine functions for absolute time encoding and combine them with relative position encoding logic for relative time encoding to generate a time vector that reflects the time-series relationship. Step B2: Normalize the numerical features such as event intensity, probability of occurrence, and intensity, and then map them to a semantic numerical vector with the same dimension as the text semantic vector through a fully connected neural network.

[0020] like Figure 2 As shown, step C specifically includes: Step C1: Load or dynamically construct the corresponding event schema from a predefined or self-learning-generated domain knowledge base. This schema defines the event type (such as "traffic congestion") and its core argument slots in a structured manner, such as time, location, intensity, and propagation probability, and specifies the data type of each slot.

[0021] Step C2: Based on their semantic information, map and fill the location slot and intensity slot in the Schema with the text semantic vector (such as the vector for "a certain intersection") generated in Step A and the numerical vector (such as the vector for "congestion intensity 0.8") generated in Step B respectively.

[0022] Step C3: Within each slot, a lightweight slot attention mechanism is employed. This mechanism uses the slot's role vector as the query and the semantic and numerical vectors assigned to that slot as keys to calculate the fusion weights, generating a slot-level fusion vector that simultaneously reflects the textual semantics and numerical features of that slot.

[0023] Step C4: Aggregate the slot-level fusion vectors generated from all slots in the schema (such as time, location, intensity, etc.). The aggregation method can be average pooling, or a schema-level attention network can be introduced to assign appropriate weights to different slots and then perform weighted summation to finally output a fixed-dimensional, unified event representation vector.

[0024] The core of the joint training and alignment process in step D lies in optimizing the consistency of the vector space, such as... Figure 3 As shown: Step D1: Construct the training dataset. Each training sample is a triple: text description, set of numerical features, and task label. The task label may include event type labels or downstream task labels, such as event priority or development trend labels. A contrastive learning method is used to construct positive sample pairs (matching text and numerical values) and negative sample pairs (non-matching pairs constructed by randomly replacing text or numerical values).

[0025] Step D2: Employ a low-rank adaptation parameter fine-tuning strategy, only updating the projection matrix of the attention layer in the pre-trained language model by injecting a trainable low-rank matrix, while simultaneously training all parameters of the numerical encoder and the schema modeling and fusion module normally.

[0026] Step D3: Define the joint loss function. This function must contain at least two parts: Contrastive learning loss aims to maximize the cosine similarity of positive sample pairs (such as the text vector of "severe congestion" and the numerical vector with an intensity value of 0.8) in the vector space, while minimizing the similarity of negative sample pairs (such as the text vector of "mild congestion" and the numerical vector with an intensity value of 0.8).

[0027] Downstream task loss: Input the unified event representation vector generated in step C4 into a classifier or regressor to predict the event type or other attributes, and calculate the cross-entropy or mean squared error loss.

[0028] Step D4: Using the backpropagation algorithm, jointly optimize all the above losses and update the parameters of the text semantic embedding module, numerical feature embedding module, and schema modeling and fusion module. This process enables the large model to learn how to generate a unified representation that satisfies both event semantics and accurately reflects numerical attributes, and that both are aligned in the vector space.

[0029] The system for unified representation of event description and numerical features based on large models includes: a text semantic embedding module with a pre-trained large language model, a numerical feature embedding module with a numerical encoder, a schema modeling and fusion module, and a joint training and alignment module.

[0030] The text semantic embedding module first receives the input natural language event description and transforms it into a token sequence through word segmentation and sub-word partitioning. Then, it automatically or according to rules identifies event trigger words and key arguments in the description and adds special tags to them. Finally, it utilizes the Transformer layer of a large language model to generate a semantic feature vector that incorporates the global context.

[0031] The numerical feature embedding module receives structured numerical inputs from databases, sensors, or other systems, such as the timestamp "2025-12-01 08:00", intensity value "0.85", and probability "0.6". This module contains multiple dedicated numerical encoders: for time features, it encodes absolute time using a sine-cosine periodic function, combined with relative time encoding; for scalars such as intensity and probability, it first normalizes them, then performs a nonlinear transformation through one or more fully connected layers, projecting them onto a semantic space that matches the dimension of the text semantic vector to form a numerical feature vector.

[0032] The Schema Modeling and Fusion Module: When processing specific events, this module selects or activates the corresponding schema based on the event type, such as a traffic congestion event schema. Then, it precisely assigns the location vector from the text semantic embedding module and the intensity vector from the numerical feature embedding module to the corresponding slots in the schema according to their semantics. Next, in the slot fusion unit, attention-weighted fusion is performed on the text and numerical vectors within each slot. Finally, the vector aggregation unit generates a unified event representation vector based on the fusion results of all slots.

[0033] Joint Training and Alignment Module: During the system training phase, this module is responsible for loading the training dataset and generating batches, while also constructing positive and negative sample pairs for contrastive learning. It implements parameter fine-tuning strategies to control the update range of the large language model parameters. The core of this module is to calculate and optimize the multi-task joint loss function, including contrastive loss and classification loss, and distribute gradients to the aforementioned three modules through backpropagation, enabling the system to learn the alignment relationship between semantics and numerical values.

[0034] Figure 2 This is a schematic diagram illustrating the structured alignment and fusion process of semantics and numerical values ​​under an event schema provided by this invention. Event schema modeling is the structured carrier for achieving a unified representation in this method. First, unstructured linguistic descriptions and discrete numerical features are mapped together into a predefined or self-learned structured schema. This schema acts as an event template, explicitly defining the type of event, such as "traffic congestion" or "earthquake," as well as a series of slots with specific semantic roles, such as time, location, intensity, and probability.

[0035] Structured alignment process: such as Figure 2 As shown, for the input event "Severe congestion occurred at Jianshe intersection in 2024, intensity 0.8", the system performs text semantic embedding and numerical semantic embedding in parallel. Subsequently, through information extraction or end-to-end learning, the text semantic vector of "Jianshe intersection" is filled into the "location" slot of the schema, the semantic vector of "severe congestion" is associated with the event type and description slots, and the time vector of "2024" and the numerical vector of "intensity 0.8" are filled into the time and intensity slots respectively. This process ensures that information is accurately placed in its semantic context and avoids confusion.

[0036] Slot-level fusion and vector aggregation: Within each slot, deep fusion is performed through a lightweight slot attention mechanism. Guided by slot role embedding, this mechanism calculates the contribution weights of text and numerical vectors to the slot representation, generating a novel slot-level vector that integrates both information. Finally, the fused vectors generated from all slots (time, location, type, intensity, etc.) in the event schema are aggregated through pooling or a schema-level attention network, outputting a unified event representation vector. This vector simultaneously encodes the event's semantics and numerical value.

[0037] To address the inherent challenges in risk event understanding tasks, the unified representation method constructed in this invention needs to be solved from multiple cognitive levels, as shown in the appendix. Figure 3 The principle of vector space alignment explained: Feature Matching Layer: For the same event, such as a traffic jam described as "traffic congestion stopped" while sensor data describes "speed reduced to 5 km / h," traditional methods struggle to directly correlate them. This invention's joint training and contrastive learning mechanism constructs positive sample pairs (matching descriptions and values) and negative sample pairs (mismatched combinations), and utilizes a contrastive loss function to bring positive sample pairs closer together and distance negative sample pairs apart in the vector space. For example... Figure 4 and Figure 5 As shown, this causes the "severely congested" text vector to be close to each other in space with high-intensity numerical vectors, while being far away from low-intensity numerical vectors.

[0038] Type Generalization Layer: The model learns a generalized event pattern schema from specific events. Different events have different specific descriptions and values, but share slots for location, time, intensity, etc. During training, events of the same type share the same schema framework, and the model learns how to fill this fixed framework with the specific descriptions and values ​​of risk events. This enables the model to abstract event structure knowledge beyond specific instances, achieving generalization from concrete to general.

[0039] Semantic and Reasoning Consistency: The model's internal representation not only matches and classifies but also supports reasoning based on numerical logic, such as "congestion with an intensity greater than 0.7 is more likely to cause traffic paralysis." Traditional pure text models perform poorly in this regard. This invention addresses this challenge through semantic embedding of numerical features and joint optimization objectives. Numerical values ​​are encoded into vectors with mathematical regularities and jointly trained with semantic vectors in a unified space. This results in the final event representation vector implicitly encoding numerical relationships such as "intensity 0.8 > intensity 0.3," thus laying the foundation for numerical-based logical reasoning.

[0040] Training Process: The training process requires a large-scale, high-quality training corpus, which should include event description text, corresponding structured numerical features, and event type or downstream task labels. Then, the joint training process is initiated. The training objective is driven by three parts of loss: 1) Contrastive learning loss: ensures that the matched semantic-numerical pairs of vectors are similar; 2) Event classification loss: Ensures that the unified vector can accurately predict event types; 3) Downstream task losses: Combined with the prediction of the impact of specific risk events.

[0041] We employ low-rank adaptation parameter fine-tuning technology to optimize some parameters of the large language model, as well as the parameters of the numerical encoder, schema modeling and fusion module, to achieve alignment between semantic and numerical spaces.

[0042] Inference Process: After model training, for unknown events, the system receives their linguistic descriptions or numerical features. Even with only partial information, the model can generate a complete event representation vector containing potential numerical relationships based on the learned unified representation space. This vector can be directly used for: 1) Event Classification and Query: Quickly determine the event type or retrieve similar events; 2) Data completion and prediction: Inferring missing numerical attributes based on known data; 3) Situation simulation and decision support: Input the event representation vector into the downstream prediction model to assess the event development trend or impact, and provide quantitative basis for intelligent traffic scheduling, emergency resource allocation, etc.

[0043] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, semiconductor memory, optical memory, etc.) containing computer-usable program code.

[0044] This invention is described with reference to flowchart or system architecture diagrams of event language descriptions and numerical feature unified correspondence representation methods, systems, and computer program products according to embodiments of the invention. It should be understood that each process or module in the flowchart and / or architecture diagram, as well as combinations of processes or modules in the flowchart or architecture diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, server, or large-scale intelligent processing platform, or other programmable data processing device to produce a machine, such that the instructions, executable by the processor of the computer or other programmable data processing device, generate instructions for implementing the process... Figure 1 One or more processes or architectures Figure 1 A device for a function specified in one or more modules.

[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer, server, or dedicated AI computing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes or architectures Figure 1 The specified functions are defined in one or more modules. Specifically, the instruction device can implement functions such as text semantic embedding, numerical feature semantic embedding, event schema modeling and fusion, and joint training and alignment.

[0046] These computer program instructions can also be loaded onto a computer, server cluster, or programmable data processing device in the cloud, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 One or more processes or architectures Figure 1 The steps of a function specified in one or more modules. These steps include, but are not limited to: semantic vectorization of input event descriptions, semantic encoding of input numerical values, structured alignment and fusion based on schema, and optimization of the unified representation through joint learning.

[0047] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. For example, adaptive adjustments or replacements can be made to the specific definition of the event schema, the structure of the numerical embedding network, the weight settings of the loss function in joint training, or the type of the underlying pre-trained large model used. Therefore, if these modifications and variations of this invention fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations. < / trigger>

Claims

1. A system for unified representation of event description and numerical features based on a large model, characterized in that, It includes a text semantic embedding module with a pre-trained large language model, a numerical feature embedding module with a numerical encoder, a schema modeling and fusion module, and a joint training and alignment module; The text semantic embedding module is used to semantically embed the natural language description of risk events, generate semantic vectors, and output them to the schema modeling and fusion module; The numerical feature embedding module is used to semantically embed the numerical features of risk events, generate numerical vectors, and output them to the schema modeling and fusion module. The Schema modeling and fusion module is used to receive semantic vectors and numerical vectors, and to perform structured alignment and fusion of the received data based on event schema modeling to generate unified event representation vectors. The joint training and alignment module establishes bidirectional data interaction with the text semantic embedding module, the numerical feature embedding module, and the schema modeling and fusion module, respectively. Through joint training and alignment, semantic vectors and numerical vectors achieve semantically consistent representation in a unified vector space.

2. The system for unified representation of event description and numerical features based on a large model according to claim 1, characterized in that, The schema modeling and fusion module includes a schema loading unit, a slot mapping unit, a slot fusion unit, and a vector aggregation unit; The Schema loading unit is used to load or dynamically build the event schema; the slot mapping unit is used to receive semantic vectors and numerical vectors and map them to the corresponding slots of the constructed schema; the slot fusion unit is used to generate a slot-level fusion vector for each slot through a lightweight slot attention mechanism; and the vector aggregation unit is used to aggregate all slot-level fusion vectors and output a unified event representation vector.

3. The system for unified representation of event description and numerical features based on a large model according to claim 1, characterized in that, The joint training and alignment module includes a dataset construction unit, a parameter fine-tuning unit, a loss calculation unit, and a parameter update unit; The dataset construction unit is used to construct the training dataset, combining natural language descriptions, numerical feature sets and task labels into sample triplets, and constructing positive sample pairs that match natural language descriptions and numerical features and negative sample pairs that randomly replace any element. The parameter fine-tuning unit is used to update some parameters of the large language model using a low-rank adaptive parameter fine-tuning strategy; The loss calculation unit is used to calculate the joint loss function, and the calculated loss result is synchronously transmitted to the parameter update unit. The parameter update unit receives the loss result from the loss calculation unit, jointly optimizes the loss function through the backpropagation algorithm, and synchronously updates various parameters of the text semantic embedding module, the numerical feature embedding module, and the schema modeling and fusion module.

4. A method for unified representation of event description and numerical features based on a large model, applied to the system described in any one of claims 1-3, characterized in that, Includes the following steps: Step A: Semantically embed the natural language description of the risk event to generate a semantic vector; Step B: Semantically embed the numerical features of the risk event to generate a numerical vector; Step C: Based on event schema modeling, the semantic vector and numerical vector are structurally aligned and fused to generate a unified event representation vector; Step D: Based on the event representation vector, through joint training and alignment, the semantic vector and the numerical vector achieve semantically consistent representation in a unified vector space.

5. The method for unified representation of event description and numerical features based on a large model according to claim 4, characterized in that, Step A specifically includes: The input natural language description is segmented and divided into sub-words, and then converted into a token sequence; Automatically or according to rules, identify event trigger words and key arguments in the token sequence and add special tags to them; The Transformer layer of a large language model is used to encode the token sequence with special tags to generate a semantic vector that incorporates the global context.

6. The method for unified representation of event description and numerical features based on a large model according to claim 4, characterized in that, Step B specifically includes: Receive structured numerical feature inputs from databases, sensors, or other systems. The numerical features include time-series features and intensity and probability features. For time-series features, a hybrid approach of absolute and relative time encoding is used to vectorize them, generating numerical vectors for time. For intensity and probability features, they are first normalized, and then nonlinearly transformed through one or more fully connected layers to project them onto a semantic space that matches the semantic vector dimension, generating numerical vectors for intensity and probability.

7. The method for unified representation of event description and numerical features based on a large model according to claim 4, characterized in that, Step C specifically includes: Step C1: Load or dynamically construct the corresponding event schema from a predefined or self-learning-generated domain knowledge base. The schema defines the event type and its core argument slots in a structured manner, and specifies the data type of each slot. Step C2: Based on the data type of the slot and the semantic information carried by the semantic vector and the numerical vector, map the semantic vector and the numerical vector to the corresponding slots in the schema respectively; Step C3: Within each slot, a lightweight slot attention mechanism is used. The role vector of the slot is used as the query, and the semantic vector and numerical vector assigned to the slot are used as the key. The fusion weight is calculated to generate a slot-level fusion vector that simultaneously reflects the natural language semantic and numerical features of the slot. Step C4: Aggregate the slot-level fusion vectors of all slots and output a unified event representation vector with fixed dimensions.

8. The method for unified representation of event description and numerical features based on a large model according to claim 7, characterized in that, In step C4, the aggregation method is average pooling, or a schema-level attention network is introduced to assign preset weights to different slots and then perform weighted summation.

9. The method for unified representation of event description and numerical features based on a large model according to claim 4, characterized in that, Step D specifically includes: Step D1: Construct a training dataset. Each training sample is a triple consisting of a natural language description, a set of numerical features, and a task label. The task label includes an event type label, an event priority label, or a development trend label. A contrastive learning method is used to construct positive sample pairs consisting of matching natural language descriptions and numerical features, as well as negative sample pairs constructed by randomly replacing natural language descriptions or numerical features. Step D2: Use a low-rank adaptation parameter fine-tuning strategy to update the projection matrix of the attention layer in the large language model by injecting a trainable low-rank matrix, while training all parameters of the numerical encoder and schema modeling and fusion module normally. Step D3: Define the joint loss function, which includes the contrastive learning loss and the downstream task loss. The downstream task loss is calculated based on the unified event representation vector output in step C. Step D4: Optimize the joint loss function using the backpropagation algorithm, and simultaneously update various parameters of the text semantic embedding module, numerical feature embedding module, and schema modeling and fusion module.

10. The method for unified representation of event description and numerical features based on a large model according to claim 4, characterized in that, The contrastive learning loss is used to constrain the cosine similarity of positive sample pairs in the vector space to approach 1, and to constrain the cosine similarity of negative sample pairs to approach 0; the downstream task loss is the cross-entropy loss or mean squared error loss calculated after inputting the unified event representation vector into the classifier or regressor.

Images