Generative event extraction method based on fusion graph and dynamic prefix tuning

Abstract

The invention relates to a generative event extraction method based on a fusion graph and dynamic prefix tuning, which comprises the following steps: S1, acquiring an input text, and carrying out structured preprocessing on the input text; s2, extracting at least two different information graphs of the structured semantic representation information, and performing linearization processing on the different information graphs to obtain corresponding sequence representations; s3, according to the sequence representations corresponding to the different information graphs, constructing prefix vector representations corresponding to the different information graphs and used for event extraction, and fusing the prefix vector representations of the different information graphs to obtain a fused prefix vector representation; s4, according to the context of the current input text, adjusting the fused prefix vector to obtain an adjusted dynamic prefix vector representation; s5, inputting a pre-training generation model, and obtaining natural language description output of the extracted target event; and S6, describing output according to the natural language so as to realize generative event extraction.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a generative event extraction method based on fusion graphs and dynamic prefix tuning. Background Technology

[0002] Event extraction is an important task in the field of natural language processing. Its goal is to identify and extract structured information related to events from unstructured text, including event type, trigger words, arguments, and their semantic roles. Effective event extraction techniques are a crucial foundation for applications such as building structured knowledge bases, information retrieval, and intelligent decision-making.

[0003] Traditional event extraction methods typically rely on manually designed rule templates or feature sets for identification and extraction. These methods are effective in specific scenarios or domains, but they often have limitations and insufficient generalization ability when dealing with texts that have complex semantic structures, frequent polysemous expressions, or diverse content.

[0004] In recent years, with the development of deep learning and pre-trained language model technologies, event extraction methods based on generative models have gradually attracted attention. These methods can extract multiple elements of an event simultaneously in an end-to-end text generation manner, reducing the complexity of feature engineering. However, existing generative event extraction methods still face certain challenges in areas such as structured semantic information fusion and dynamic modeling of contextual semantic structures. Especially when dealing with complex semantic scenarios or the coexistence of multiple events, the model may exhibit problems such as missing or inaccurate generated structural information. Summary of the Invention

[0005] This invention provides a generative event extraction method based on fusion graphs and dynamic prefix tuning. By using artificial intelligence technology and combining complementary semantic structure information provided by information graphs with various forms of structured semantic representation information, prefix representations are dynamically constructed to assist generative models in better capturing the structural features of events, providing effective support for tasks such as event type identification, trigger word localization, and event argument extraction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, this application provides a generative event extraction method based on fusion graphs and dynamic prefix tuning, including: S1, Obtain the input text and perform structured preprocessing on the input text to obtain the preprocessed text; S2, based on the preprocessed text, extract information graphs of at least two different structured semantic representations, and perform linearization processing on the different information graphs to obtain the corresponding sequence representations; S3. Based on the sequence representations corresponding to different information graphs, construct the prefix vector representations for event extraction corresponding to different information graphs, and fuse the prefix vector representations of different information graphs to obtain the fused prefix vector representation. S4. Adjust the fused prefix vector according to the context of the current input text to obtain the adjusted dynamic prefix vector representation; S5, input the dynamic prefix vector representation into the pre-trained generation model to obtain the natural language description output of the extracted target event; S6, based on the natural language description output, obtains the event type, event trigger words, event arguments and their corresponding role information in the text through template parsing, so as to achieve generative event extraction.

[0007] In one implementation, in S2, the information graph includes an abstract meaning representation graph (AMR) and an information extraction graph (IE).

[0008] In one implementation, an abstract semantic representation parsing tool is used to parse the preprocessed text into an AMR graph structure; an open information extraction tool is used to extract the IE graph from the text.

[0009] In one implementation, in S3, the sequence representations of different information graphs are semantically encoded using the corresponding Transformer encoders. The structured semantic information in the encoding results is extracted using a multi-head attention mechanism and fused to obtain a dynamic prefix vector representation adapted to the current input text.

[0010] In one implementation, the fusion strategy is weighted fusion.

[0011] In one implementation, in step S4, a context-aware prefix gating mechanism is used to dynamically evaluate the semantic relevance between the prefix representation and the current input context, and a diversity-enhanced prefix regularization strategy is adopted to adjust the fused prefix vector representation to obtain the adjusted dynamic prefix vector representation.

[0012] In one implementation, the pre-trained model integrates a dynamic replication mechanism.

[0013] In one implementation, the process of generating a natural language description of an event by the model may involve two modes simultaneously: generating new words from the vocabulary and copying entities from the original input text.

[0014] The technical solution of this invention utilizes Abstract Meaning Representation Graph (AMR) and Information Extraction Graph (IE) as external sources of structured semantic information, and combines them with dynamic prefix optimization technology to more accurately capture the structured semantic features in the text, thereby improving the accuracy and reliability of identifying event types, locating event trigger words, and extracting event arguments and their roles from unstructured text. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall model framework of an embodiment of the present invention; Figure 2 A diagram illustrating the context-aware prefix gating (CPG) mechanism; Figure 3 This is a diagram illustrating the combined effects of prefix regularization and dynamic replication mechanisms. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0017] To address the shortcomings and problems of existing technologies, this application provides a generative event extraction method based on fusion graphs and dynamic prefix tuning, comprising: S1, Obtain the input text and perform structured preprocessing on the input text to obtain the preprocessed text; S2, based on the preprocessed text, extract information graphs of at least two different structured semantic representations, and perform linearization processing on the different information graphs to obtain the corresponding sequence representations; S3. Based on the sequence representations corresponding to different information graphs, construct the prefix vector representations for event extraction corresponding to different information graphs, and fuse the prefix vector representations of different information graphs to obtain the fused prefix vector representation. S4. Adjust the fused prefix vector according to the context of the current input text to obtain the adjusted dynamic prefix vector representation; S5, input the dynamic prefix vector representation into the pre-trained generation model to obtain the natural language description output of the extracted target event; S6, based on the natural language description output, obtains the event type, event trigger words, event arguments and their corresponding role information in the text through template parsing, so as to achieve generative event extraction.

[0018] Detailed Implementation Examples In a more detailed embodiment below, combined with Figures 1 to 3 The specific details of the scheme in this application shall be explained.

[0019] The method of the present invention specifically includes the following steps: (1) The input text is preprocessed in a structured manner. The abstract meaning representation graph parsing tool and open information extraction tool are used to construct the AMR graph and IE graph of the text respectively. The two types of graph structures are linearized to obtain the sequence representation that the model can input. Starting with natural language text input, the text to be analyzed is fed into the text structuring preprocessing module. This module is implemented as follows: using the Abstract Semantic Representation Parsing Tool (Amrlib), the input text is parsed into an AMR graph structure to extract core semantic features. Simultaneously, using the Open Information Extraction (OpenIE) tool, entity-relationship triples are extracted from the same input text to construct an IE graph, clarifying the details of entities and their relationships. To facilitate subsequent model processing, both the AMR graph and the IE graph are linearized into sequence forms. The AMR graph is converted into a linearized sequence through depth-first traversal, while the IE graph is formed into a linearized text sequence representation by simply concatenating triples (head entity-relationship-tail entity). Figure 1 .

[0020] (2) Sequence representations of the two types of graphs above are semantically encoded using dedicated Transformer encoders. The structured semantic information in the encoding results is extracted using a multi-head attention mechanism and fused to obtain a dynamic prefix vector representation that is adapted to the current input text. Specifically, semantic encoding modules designed for AMR and IE graph features encode semantic features for the aforementioned linearized sequences. Specifically, the linearized sequence of the AMR graph is encoded using an AMRBart encoder specifically designed for AMR to capture the abstract semantic relational structure of the sentence. The linearized sequence of the IE graph uses a standard BART encoder, focusing on capturing explicit entity and relational information in the text. Through these encoding modules, semantic feature representations corresponding to the AMR and IE graphs are obtained respectively. Figure 1 .

[0021] To make the model more sensitive to structured semantic features in the input text, this invention designs a dynamic prefix vector generation module to achieve dynamic fusion of structured semantic features. This module dynamically extracts representative semantic information from the semantic feature representations of AMR and IE through a multi-head attention mechanism, generating AMR prefix and IE prefix representations respectively. Then, through a weighted fusion strategy, the AMR prefix and IE prefix are dynamically combined into a unified dynamic prefix vector representation, thereby achieving efficient expression of the structured features of the text.

[0022] (3) Through a context-aware prefix gating mechanism, the semantic relevance between the prefix representation and the current input context is dynamically evaluated, and a diversity-enhanced prefix regularization strategy is adopted to reduce the redundancy of prefix information and improve the adaptability of the prefix representation to the input text. To enhance the adaptability of dynamic prefix vectors to input context and improve representation diversity, this invention proposes a context-aware prefix gating mechanism and a diversity-enhancing regularization strategy, such as... Figure 2 The specific implementation method is as follows: After dynamically generating prefix vectors, the semantic relevance weights of the prefix vectors are calculated based on the current input context, and the prefix vectors are dynamically filtered and adjusted to reduce redundant or irrelevant prefix information and enhance the focus on key semantics. In addition, during model training, diversity-enhancing regularization is used to reduce redundancy and similarity between prefix vectors to ensure that each prefix can effectively capture semantic features at different levels or aspects.

[0023] (4) The dynamically fused and gated prefix representation is injected layer by layer into each layer of the pre-trained generative model BART, and the natural language description output of the target event is obtained by using the generation strategy based on event prompt template. The present invention also designs a dynamic replication mechanism to further improve the stability and accuracy of the generated model.

[0024] The specific implementation involves two modes simultaneously: generating new words from the vocabulary and copying entities from the original input text during the model's generation of natural language descriptions of events. This invention utilizes a dynamic copying mechanism to dynamically adjust the ratio between generating from the vocabulary and copying entities from the input text in real time, based on the semantic matching degree between the current generation context and the input text. This avoids the problems of inaccurate generation or excessive copying inherent in traditional static copying mechanisms. This mechanism allows the model to flexibly select a more suitable generation mode based on the context, thereby effectively improving the accuracy and coherence of generated event information, especially in text scenarios with complex event structures or a large number of entities. The effects of prefix regularization and the dynamic copying mechanism are evident. Figure 3 .

[0025] (5) Based on the generated natural language description of the event, the event type, event trigger words and event arguments and their corresponding role information in the text are obtained through template parsing, so as to realize structured event extraction.

[0026] The dynamically adjusted and optimized prefix vectors are injected layer by layer into each layer of the pre-trained BART generative model to achieve a deep perception and understanding of the input semantic structure. During event generation, this invention employs natural language prompt templates tailored to specific event types. These templates define the meaning of the event type, possible trigger words, and the structural relationships of corresponding argument roles. In conjunction with the dynamic prefix mechanism, the model can accurately generate compliant and easily parsed event description text during the generation phase.

[0027] Through a specially designed template parsing module, structured event information is automatically extracted from the natural language descriptions generated by the model. This includes identifying event types, accurately locating event trigger words, and further extracting arguments and their roles within the event. Finally, a complete and clear structured event extraction result is output for subsequent application in tasks such as information retrieval, knowledge graph construction, and public opinion analysis.

[0028] In the several embodiments provided by this invention, it should be understood that the disclosed methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0029] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0030] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A generative event extraction method based on fusion graph and dynamic prefix tuning, characterized in that, include: S1, Obtain the input text and perform structured preprocessing on the input text to obtain the preprocessed text; S2, based on the preprocessed text, extract information graphs of at least two different structured semantic representations, and perform linearization processing on the different information graphs to obtain the corresponding sequence representations; S3. Based on the sequence representations corresponding to different information graphs, construct the prefix vector representations for event extraction corresponding to different information graphs, and fuse the prefix vector representations of different information graphs to obtain the fused prefix vector representation. S4. Adjust the fused prefix vector according to the context of the current input text to obtain the adjusted dynamic prefix vector representation; S5, input the dynamic prefix vector representation into the pre-trained generation model to obtain the natural language description output of the extracted target event; S6, based on the natural language description output, obtains the event type, event trigger words, event arguments and their corresponding role information in the text through template parsing, so as to achieve generative event extraction.

2. The generative event extraction method based on fusion graph and dynamic prefix tuning according to claim 1, characterized in that, In S2, the information graph includes the Abstract Meaning Representation Graph (AMR) and the Information Extraction Graph (IE).

3. The generative event extraction method based on fusion graph and dynamic prefix tuning according to claim 2, characterized in that, Using an abstract semantic representation parsing tool, the preprocessed text is parsed into an AMR graph structure; an open information extraction tool is used to extract IE graphs from the text.

4. The generative event extraction method based on fusion graph and dynamic prefix tuning according to claim 3, characterized in that, In S3, the sequence representations of different information graphs are semantically encoded using the corresponding Transformer encoders. The multi-head attention mechanism is used to extract the structured semantic information in the encoding results and fuse them to obtain a dynamic prefix vector representation that adapts to the current input text.

5. The generative event extraction method based on fusion graph and dynamic prefix tuning according to claim 4, characterized in that, The fusion strategy is weighted fusion.

6. The generative event extraction method based on fusion graph and dynamic prefix tuning according to claim 5, characterized in that, In S4, the semantic relevance between the prefix representation and the current input context is dynamically evaluated through a context-aware prefix gating mechanism, and a diversity-enhanced prefix regularization strategy is adopted to adjust the fused prefix vector representation to obtain the adjusted dynamic prefix vector representation.

7. The generative event extraction method based on fusion graph and dynamic prefix tuning according to claim 1, characterized in that, The pre-trained model integrates a dynamic replication mechanism.

8. The generative event extraction method based on fusion graph and dynamic prefix tuning according to claim 7, characterized in that, In the process of generating natural language descriptions of events using a model, two modes may be involved simultaneously: generating new words from a vocabulary and copying entities from the original input text.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a program, and the processor executing the program to implement the method of any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The device contains a program that is executed by a processor to implement the method of any one of claims 1 to 8.

Images