CN-121998095-A - System and method for uniformly representing event description and numerical characteristics based on large model

CN121998095ACN 121998095 ACN121998095 ACN 121998095ACN-121998095-A

Abstract

The invention provides a system and a method for uniformly representing event description and numerical characteristics based on a large model, which belong to the technical fields of artificial intelligence, natural language processing and large models and are used for solving the problems that event semantics and numerical characteristic representation are split and cross-modal alignment and uniform understanding are difficult to realize in the traditional technology; the method comprises the steps of A, carrying out semantic embedding on numerical characteristics of a risk event to generate numerical vector representation, B, carrying out alignment and fusion on the semantic vector and the numerical vector according to structural slots based on risk event Schema modeling, and D, carrying out unified representation of the semantic vector and the numerical vector in a unified vector space through joint training and contrast learning.

Inventors

CHANG MENGMENG
YANG YANZHI
LU YAYAN
HAN LIQIN
LI GUANGCHAO
GAO LUWEI

Assignees

河南师范大学

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (10)

1. The system for uniformly representing the event description and the numerical characteristics based on the large model is characterized by comprising a text semantic embedding module provided with a pre-training large language model, a numerical characteristic embedding module provided with a numerical encoder, a Schema modeling and fusion module and a joint training and alignment module; The text semantic embedding module is used for carrying out semantic embedding on natural language description of the risk event, generating a semantic vector and outputting the semantic vector to the Schema modeling and fusion module; the numerical feature embedding module is used for semantically embedding the numerical features of the risk event, generating a numerical vector and outputting the numerical vector to the Schema modeling and fusion module; the Schema modeling and fusing module is used for receiving the semantic vector and the numerical vector, and carrying out structural alignment and fusion on the received data based on event Schema modeling to generate a unified event characterization vector; The combined training and aligning module respectively establishes two-way data interaction with the text semantic embedding module, the numerical feature embedding module and the Schema modeling and fusing module, and semantic consistent representation is realized on semantic vectors and numerical vectors in a unified vector space through combined training and aligning.
2. The system for uniformly representing event descriptions and numerical features based on a large model according to claim 1, wherein the Schema modeling and fusion module comprises a Schema loading unit, a slot mapping unit, a slot fusion unit and a vector aggregation unit; The system comprises a Schema loading unit, a slot mapping unit, a slot fusion unit and a vector aggregation unit, wherein the Schema loading unit is used for loading or dynamically constructing an event Schema, the slot mapping unit is used for receiving semantic vectors and numerical vectors and mapping the semantic vectors and the numerical vectors to corresponding slots of the constructed Schema, the slot fusion unit is used for generating slot level fusion vectors of each slot through a lightweight slot attention mechanism, and the vector aggregation unit is used for aggregating all slot level fusion vectors and outputting unified event characterization vectors.
3. The system for unified representation of event descriptions and numerical features based on large models according to claim 1, wherein the joint training and alignment module comprises a data set construction unit, a parameter fine tuning unit, a loss calculation unit, and a parameter updating unit; The data set construction unit is used for constructing a training data set, combining the natural language description, the numerical value feature set and the task tag into a sample triplet, and constructing a positive sample pair with the natural language description matched with the numerical value feature and a negative sample pair with any element replaced randomly; the parameter fine adjustment unit is used for updating partial parameters of the large language model by adopting a low-rank adaptive parameter fine adjustment strategy; The loss calculation unit is used for calculating a joint loss function, and the calculated loss result is synchronously transmitted to the parameter updating unit; the parameter updating unit is used for receiving the loss result transmitted by the loss calculating unit, jointly optimizing the loss function through a back propagation algorithm, and synchronously updating various parameters of the text semantic embedding module, the numerical characteristic embedding module and the Schema modeling and fusion module.
4. A method for unified representation of event descriptions and numerical features based on a large model, applied to a system as claimed in any one of claims 1 to 3, comprising the steps of: step A, carrying out semantic embedding on natural language description of a risk event to generate a semantic vector; step B, semantically embedding the numerical characteristics of the risk event to generate a numerical vector; step C, based on event Schema modeling, carrying out structural alignment and fusion on the semantic vector and the numerical vector to generate a unified event characterization vector; And D, based on the event characterization vector, realizing semantic consistency representation of the semantic vector and the numerical vector in a unified vector space through joint training and alignment.
5. The method for unified representation of large model-based event descriptions and numerical features according to claim 4, wherein step a specifically comprises: dividing the input natural language description into word and sub-word, and converting the word and sub-word into Token sequence; automatically or according to rules, identifying event trigger words and key arguments in the Token sequence, and adding special marks for the event trigger words and the key arguments; and encoding the Token sequence added with the special mark by utilizing a transducer layer of the large language model to generate a semantic vector fused with the global context.
6. The method for unified representation of large model-based event descriptions and numerical features according to claim 4, wherein step B specifically comprises: Receiving structured numerical feature inputs from a database, sensor, or other system, the numerical features including time-series class features and intensity and probability class features; the method comprises the steps of carrying out vectorization on time sequence characteristics by adopting a mixed mode of absolute time coding and relative time coding to generate a numerical vector of a time sequence, carrying out normalization processing on strength and probability characteristics, carrying out nonlinear transformation through one or more full-connection layers, projecting the strength and probability characteristics to a semantic space matched with semantic vector dimensions, and generating the numerical vector of the strength and probability.
7. The method for unified representation of large model-based event descriptions and numerical features of claim 4 wherein step C specifically comprises: Step C1, loading or dynamically constructing a corresponding event Schema from a predefined or self-learning domain knowledge base, wherein the Schema defines event types and core argument slots thereof in a structural mode, and prescribes the data types of each slot; Step C2, according to the data type of the slot, semantic vectors and semantic information carried by numerical vectors, respectively mapping the semantic vectors and the numerical vectors to the corresponding slots of the Schema; Step C3, adopting a lightweight slot attention mechanism in each slot, taking a role vector of the slot as a Query, taking a semantic vector and a numerical vector distributed to the slot as key values, calculating fusion weights, and generating a slot level fusion vector which simultaneously reflects the natural language semantics and the numerical characteristics of the slot; and step C4, aggregating the slot level fusion vectors of all slots, and outputting a unified event characterization vector with fixed dimension.
8. The method for uniformly representing event descriptions and numerical features based on large models according to claim 7, wherein in step C4, the aggregation mode is average pooling or the weighted summation is performed after the predetermined weights are assigned to different slots by introducing a Schema level attention network.
9. The method for unified representation of large model-based event descriptions and numerical features of claim 4 wherein step D specifically comprises: Step D1, constructing a training data set, wherein each training sample is a triplet formed by natural language description, a numerical value feature set and a task tag, the task tag comprises an event type tag, an event priority tag or a development trend tag, a positive sample pair formed by matched natural language description and numerical value features is constructed by adopting a contrast learning method, and a negative sample pair formed by randomly replacing the natural language description or the numerical value feature is constructed; Step D2, updating a trainable low-rank matrix injected into a projection matrix of an attention layer in a large language model by adopting a low-rank adaptive parameter fine tuning strategy, and simultaneously normally training all parameters of a numerical encoder and a Schema modeling and fusion module; Step D3, defining a joint loss function, wherein the joint loss function comprises comparison learning loss and downstream task loss, and the downstream task loss is calculated based on the unified event characterization vector output by the step C; and D4, jointly optimizing a joint loss function through a back propagation algorithm, and synchronously updating various parameters of a text semantic embedding module, a numerical characteristic embedding module and a Schema modeling and fusion module.
10. The method for uniformly representing event descriptions and numerical features based on a large model according to claim 4, wherein the contrast learning loss is used for constraining cosine similarity of positive sample pairs in a vector space to be close to 1 and constraining cosine similarity of negative sample pairs to be close to 0, and the downstream task loss is cross entropy loss or mean square error loss calculated after the uniform event characterization vector is input into a classifier or a regressor.

Description

System and method for uniformly representing event description and numerical characteristics based on large model Technical Field The invention belongs to the technical fields of artificial intelligence, natural language processing and large models, and particularly relates to a system and a method for uniformly representing event description and numerical characteristics based on a large model. Background With the widespread use of large models (Large Language Models, LLMs) in tasks such as natural language understanding, event analysis and prediction, it becomes a key challenge during a risk event how to make the model understand the semantic description of an event and related numerical features (e.g., time, intensity, probability, etc.) at the same time. The description of the risk event is mainly directed to the description of a dynamic process, and the risk event contains abundant text information (such as 'serious congestion at an intersection of 2025, 12 and 22 days') and is accompanied by specific numerical attributes (such as congestion starting time, duration, influence range, traffic flow change rate and the like). Particularly, in the scenes of smart cities, emergency management, natural disaster analysis and the like, unified understanding of event semantics and numerical values can be realized, and the accuracy of system research and judgment, the timeliness of response and the interpretability of decisions are directly related. In this context, the use of large models to implement multidimensional characterization and reasoning of risk events is a research hotspot. Some scientific institutions and teams have made preliminary progress in this direction. For example, in iTransformer model proposed by the university of Qinghua software academy of sciences, the attention mechanism of the transducer is reconstructed, the correlation among the multiple variables is effectively captured through inversion processing of the time sequence dimension, and good performance is achieved in the prediction task of the numerical sequence. In EventRAG research on Zhejiang university and ant group, an enhancement generation framework based on event knowledge graph is provided. By converting unstructured documents into event map nodes containing time dependence and logic relations and utilizing agents to conduct iterative search and reasoning, the logic consistency and long Cheng Tuili capacity of the large model in complex narrative scenes are remarkably improved. However, the existing method still has obvious limitations in realizing the unified representation of event description and numerical characteristics, and common technical paths and defects thereof are as follows: (1) A text semantic embedding method; And directly encoding the event description text by using the pre-training language model to generate a semantic vector. The method treats numerical features (e.g., "30 minutes" duration) as plain text token. The method has the defects that mathematical characteristics and semantics of numerical values cannot be specially modeled, so that the correlation between 30 minutes and serious congestion in a vector space is weak, and downstream tasks based on numerical value comparison, quantitative reasoning and the like are difficult to support. (2) Post-fusion multimodal methods; the text is processed by using a language model, the numerical value is processed by using a special encoder, and after two types of vectors are obtained, fusion is carried out by vector splicing, weighted addition or a simple attention mechanism. The method has the defects that the text and the numerical vector are usually independently generated in the initial stage of training, lack of explicit constraint of cross-modal alignment, cause 'semantic-numerical disjoint', and describe that the text vector of 'congestion aggravation' and the numerical vector of 'speed drop rate 0.8' are far away in space, so that the consistency of the overall event characterization is affected. (3) Rule or template driven structuring methods; And predefining a Schema for the event, including fixed slots of time, place, intensity and the like, and filling texts and numerical values into the corresponding slots through an information extraction technology to form a structured record. The method has the defects that the method depends on manual design Schema, has poor generalization capability, only realizes alignment of a symbol layer, does not realize deep fusion of semantics and numerical values on a vector representation layer, and is difficult to support complex reasoning based on representation. Although research on understanding of events is advanced in different layers by the existing method, the core problems of weak semantic and numerical representation fracture and alignment mechanism and unified representation of frame deletion generally exist. The large model is limited in performance when dealing with complex tasks of what and how what happe