A hybrid prompt generation method and device for document-level event argument extraction

Abstract

The application relates to the technical field of event extraction, and discloses a hybrid prompt generation method and device for document-level event argument extraction. The method realizes accurate constraint through fine-grained concept guidance prompts, realizes dynamic enhancement through soft instance perception memory, realizes synergistic fusion through a hybrid generator, and constructs an efficient, accurate and robust hybrid prompt generation model. Based on the hybrid prompt generation model, the coverage of document-level event argument extraction and the effectiveness of retrieval enhancement can be improved, and more comprehensive event argument extraction can be realized.

Description

Technical Field

[0001] This invention relates to the field of event extraction technology, and in particular to a hybrid prompt generation method and apparatus for document-level event argument extraction. Background Technology

[0002] Document-level Event Argument Extraction (DocEAE) aims to identify event arguments and their corresponding roles from the entire document, thereby transforming unstructured event information into a structured representation of event arguments. By extracting event arguments scattered across multiple sentences, DoCEEA supports a range of downstream tasks, including information retrieval, question answering, and dialogue systems. Compared to traditional event-level event argument extraction, DoCEEA faces additional challenges stemming from long-distance dependencies between event trigger words and arguments, as well as the frequent co-occurrence of multiple events within a single document. These factors significantly increase the complexity of the task and limit the performance of traditional event argument extraction methods.

[0003] Existing DoCEAE methods are generally divided into two main categories: classification-based methods and generative methods. Classification-based methods formulate event argument extraction as a discriminative classification task, assigning role labels to candidate spans to produce structured event output. Specifically, classification-based methods typically utilize contextual representations to classify candidate text spans, thereby identifying argument boundaries and assigning their respective semantic roles. For example, combining encoding and AMR graphs collaboratively extracts semantic features from local and global dimensions for span classification and extraction. Argument extraction can also be transformed into a prototype-driven classification problem, which uses hyperspheres and multi-prototype optimal transport matching to improve extraction accuracy. Driven by the success of Large Language Models (LLMs), generative methods have become the dominant paradigm. By redefining DoCEAE as a conditional text generation task, these methods rely on generative models to extract event arguments. For example, an end-to-end framework performs document-level event argument extraction under the guidance of pre-set templates through conditional generation. Furthermore, DOCEAE can be transformed into conditional text generation through semantically enhanced cues, which guide generative models to extract event arguments within a unified framework. To further improve the performance of generative models in DOCEAE, recent research has introduced cue enhancement strategies to incorporate richer prior information. For example, learnable soft cues are introduced, allowing cue representations to adapt to different examples, thus enabling more effective integration of relevant contextual information. Additionally, there are methods combining joint multi-view unanswered cues with ontology-driven decoding, injecting complementary cross-cue evidence and event-specific knowledge constraints into the generation process. Other methods introduce a reasoning-driven cue strategy, enabling LLMs to explicitly extract task reasoning from demonstrations for cue enhancement, thereby improving complex argument extraction in DOCEAE. While current cue enhancement methods have achieved impressive performance on multiple benchmark datasets, they struggle to effectively guide models to focus on key information when dealing with DOCEAE tasks, resulting in limited extraction efficiency.

[0004] The success of Retrieval-Augmented Generation (RAG) in natural language processing has spurred its application in event argument extraction. By introducing a document-level dynamic memory module to store contextual event information, the extraction of long-distance dependent arguments is enhanced. Another approach proposes a DoCEAE-based framework that ranks events by difficulty and leverages high-confidence events to build a reliable memory for predicting more complex events. RAG supplements prior information in the input by retrieving external knowledge, thus providing better guidance for the model's generation process. However, these methods largely ignore two key limitations: input length constraints and insufficient retrieval diversity, which limit the coverage of context and the effectiveness of retrieval enhancement.

[0005] It is evident that existing document-level event argument extraction methods suffer from narrow coverage and poor effectiveness in enhancing retrieval. Summary of the Invention

[0006] Based on this, this application provides a hybrid prompt generation method and apparatus for document-level event argument extraction to solve the problems existing in the prior art.

[0007] Firstly, this application provides a hybrid prompt generation method for document-level event argument extraction, the method comprising:

[0008] Design fine-grained concept guidance prompts, which include event type, event trigger words and location information, as well as argument roles;

[0009] A soft instance-aware memory is constructed to store representations of all training instances and selectively retrieve information. The soft instance-aware memory includes memory construction, memory retrieval, and adaptive fusion mechanisms.

[0010] A hybrid generator is built based on the decoder, which is used to generate arguments by fusing fine-grained concept guidance prompts and soft instance-aware memory;

[0011] A hybrid prompt generation model is constructed based on fine-grained concept guidance prompts, soft instance-aware memory, and a hybrid generator;

[0012] Argument extraction based on a hybrid prompting generation model and real-time documents.

[0013] In one embodiment, the event type is used to provide a fine-grained semantic description of the event type; the event trigger word and location information are used to specify the trigger word and its corresponding position in the input document; and the argument roles are used to provide a natural language template with role placeholders.

[0014] In one embodiment, fine-grained concept guidance prompts satisfy the following relationship:

[0015]

[0016] in, This indicates a splicing operation. It is an event type guideline. It is a trigger word guide prompt. It is an argument role guide prompt;

[0017] The method also includes:

[0018] Input document With corresponding fine-grained concept guidance tips Connect to form the encoder input, as follows:

[0019]

[0020] The contextual representation of the input is computed using a multi-layer Transformer encoder. ,as follows:

[0021]

[0022] in , Indicates the maximum sequence length of the encoder. Indicates the hidden dimension. Each layer is implemented as a Transformer encoder block with a self-attention mechanism.

[0023] In one embodiment, the construction process of the memory structure is as follows:

[0024] Given a training corpus ,Depend on It consists of documents, each document... The input format is converted to a unified form as follows:

[0025]

[0026] In the formula, , Refers to the start and end characters;

[0027] The input is fed into the shared encoder to obtain the start character. Context representation It is used as an instance representation. ,as follows:

[0028]

[0029] in All instances together form a soft, instance-based memory. ,as follows:

[0030] .

[0031] In one embodiment, the memory retrieval process is constructed as follows:

[0032] Instance-based memory and the context representation of the input ,exist Max pooling is applied to obtain a global semantic representation. It is used for querying, as follows:

[0033]

[0034] in ;

[0035] Using cosine similarity Measuring the correlation between the input sequence and the soft instances stored in the instance store, using queries from Before searching The most relevant instances are represented as explicit learning of relevant events, as follows:

[0036]

[0037] In the formula, These are the top K most relevant instances retrieved.

[0038] In one embodiment, the adaptive fusion mechanism is used to construct an aggregated instance representation, and the construction process is as follows:

[0039] Query The similarity score between each retrieved instance representation is converted into a normalized attention weight. To generate memory output, where the query The cosine similarity S between each retrieved instance representation and the following calculation:

[0040]

[0041] In the formula, Represents the relationship between the i-th vector and the query. Cosine similarity between them ;

[0042] The relevant attention weights are defined as:

[0043]

[0044] In the formula, Represents the relationship between the j-th vector and the query. Cosine similarity between them , It's a hyperparameter. These are the top K most relevant instances retrieved;

[0045] Computational aggregate memory representation The weighted sum, representing the retrieved instances, is as follows:

[0046]

[0047] In the formula, .

[0048] In one embodiment, building a hybrid generator based on the decoder includes:

[0049] The computationally enhanced encoder is represented as:

[0050]

[0051] In the formula, and Represents trainable parameters. ;

[0052] In the generation step Multi-layer Transformer decoder predicts the first Output token and decoder hidden state as follows:

[0053]

[0054] in, Each layer is implemented as a Transformer decoder block, which includes the previously generated decoder state. Masking self-attention and representation of the augmented encoder Cross attention;

[0055] The conditional probability of the entire output sequence is decomposed by autoregression as follows:

[0056]

[0057] In the formula, This indicates the number of terms in the output sequence.

[0058] In one embodiment, the learning objective of the hybrid prompt generation model is to minimize the negative log-likelihood loss:

[0059]

[0060] In the formula, Let be the conditional probability of the entire output sequence. This is the training dataset.

[0061] Secondly, this application provides a hybrid prompt generation apparatus for document-level event argument extraction, comprising:

[0062] The first unit is used to design fine-grained concept guidance prompts, which include event types, event trigger words and location information, as well as argument roles;

[0063] The second unit is used to construct soft instance-aware memory to store representations of all training instances and selectively retrieve information. Soft instance-aware memory includes memory construction, memory retrieval, and adaptive fusion mechanisms.

[0064] The third unit is used to build a hybrid generator based on the decoder. The hybrid generator is used to generate arguments by fusing fine-grained concept guidance prompts and soft instance-aware memory.

[0065] The fourth unit is used to build a hybrid prompt generation model based on fine-grained concept guidance prompts, soft instance-aware memory, and hybrid generators;

[0066] The fifth unit is used for argument extraction based on a hybrid prompting generation model and real-time documents.

[0067] Thirdly, this application provides a hybrid prompt generation apparatus for document-level event argument extraction, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.

[0068] Beneficial effects:

[0069] The hybrid prompt generation method for document-level event argument extraction provided by this invention achieves precise constraints through fine-grained concept-guided prompts, achieves dynamic enhancement through soft instance-aware memory, and achieves collaborative fusion through a hybrid generator. This constructs an efficient, accurate, and robust hybrid prompt generation model. Based on this hybrid prompt generation model, the coverage of document-level event argument extraction and the effectiveness of retrieval enhancement can be improved, enabling more comprehensive event argument extraction. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a flowchart illustrating a hybrid prompt generation method for document-level event argument extraction in one embodiment;

[0072] Figure 2 This is a schematic diagram of the overall framework of a hybrid prompt generation method for document-level event argument extraction in one embodiment. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0074] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0075] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0076] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0077] This application provides a hybrid prompt generation method for document-level event argument extraction, the method comprising:

[0078] Design fine-grained concept guidance prompts, which include event type, event trigger words and location information, as well as argument roles.

[0079] In this step, to accurately identify the event arguments for each event in the document, a fine-grained concept guidance prompt was designed, containing three levels of conceptual information: event type, event trigger word, and argument role. To further alleviate semantic confusion caused by the coexistence of multiple events, the positional information of each event trigger word is explicitly embedded in the prompt. In this way, precise constraints can be achieved through the fine-grained concept guidance prompt.

[0080] A soft instance-aware memory is constructed to store representations of all training instances and selectively retrieve information. The soft instance-aware memory includes memory construction, memory retrieval, and an adaptive fusion mechanism.

[0081] In this step, to address the limitations of existing models on maximum input length and insufficient diversity of retrieved instances, a soft instance memory is proposed. This soft instance memory stores representations of all training instances and selectively retrieves information, which can facilitate explicit learning of relevant events.

[0082] A hybrid generator is built based on the decoder. The hybrid generator is used to generate arguments by fusing fine-grained concept guidance prompts and soft instance-aware memory.

[0083] In this step, to fully utilize both abstract conceptual information and concrete instance information, a hybrid enhancement mechanism is proposed to improve the encoder representation by integrating retrieved instance information. This process incorporates instance-level contextual information while preserving the original semantic content, thus enabling better collaborative fusion.

[0084] A hybrid prompt generation model is constructed based on fine-grained concept guidance prompts, soft instance-aware memory, and a hybrid generator.

[0085] Argument extraction based on a hybrid prompting generation model and real-time documents.

[0086] The aforementioned hybrid prompt generation method for document-level event argument extraction achieves precise constraints through fine-grained concept-guided prompts, dynamic enhancement through soft instance-aware memory, and collaborative fusion through a hybrid generator. This constructs an efficient, accurate, and robust hybrid prompt generation model. Based on this hybrid prompt generation model, the coverage of document-level event argument extraction and the effectiveness of retrieval enhancement can be improved, enabling more comprehensive event argument extraction.

[0087] The following is a complete example illustrating the hybrid prompt generation method described above for document-level event argument extraction:

[0088] 1. The following are guiding tips for designing fine-grained concepts:

[0089] To accurately identify the event arguments of each event in a document, this application designs a fine-grained concept guidance prompt, which contains three levels of concept information: event type, event trigger word, and argument role. To further alleviate semantic confusion caused by the coexistence of multiple events, this application explicitly embeds the positional information of each event trigger word into the prompt. For example... Figure 2 As shown, the fine-grained concept guidance consists of the following components:

[0090] Event type definition It provides a fine-grained semantic description of event types. This explicit semantic guidance helps the model better understand the nature of events and improves the accuracy of event argument extraction.

[0091] Trigger words and location information It specifies the event trigger word and its corresponding position in the input document, enabling the model to more accurately locate the trigger word. This explicit positional description enhances the model's ability to focus on the correct span.

[0092] Event Argument Template with Character Placeholders It provides natural language templates with role placeholders, such as "someone," "an organization," and "somewhere," guiding the model to generate complete and fluent argument expressions.

[0093] Based on the above components, fine-grained concept guidance prompts are defined as follows:

[0094]

[0095] in, This indicates a concatenation operation, and the symbol must be followed thereafter.

[0096] Input document With corresponding fine-grained concept guidance tips Connect to form encoder input.

[0097]

[0098] To this end, HyGEN first computes the contextual representation of the input through a multi-layer Transformer encoder. .

[0099]

[0100] in , Indicates the maximum sequence length of the encoder. This represents the hidden dimension. Each layer is implemented as a Transformer encoder block with a self-attention mechanism.

[0101] 2. The soft instance-aware memory is constructed as follows:

[0102] First, it is worth noting that although RAG performs well in event argument extraction, it has two key limitations: (1) the limitation on the maximum input length of the model, and (2) the lack of diversity in retrieved instances. To address these issues, this application proposes a soft instance memory that stores representations of all training instances and selectively retrieves information to facilitate explicit learning of relevant events.

[0103] The soft instance-aware memory consists of three parts: memory construction, memory retrieval, and adaptive fusion mechanism. The specific implementation details are as follows.

[0104] Memory construction. Given a training corpus. ,Depend on It consists of 1 document, each document It will be converted into a unified input format. Then, It will be fed into the shared encoder to obtain a special start token. Context representation It is used as an instance representation. ,as follows:

[0105]

[0106] in All instances together form a soft, instance-based memory. ,as follows:

[0107] .

[0108] It is important to note that the soft instance-level memory is entirely constructed from the current training corpus. The memory stores the vector representation of each instance, which is also called a soft instance.

[0109] Memory retrieval. Instance-based memory. and input representation This application is in Max pooling is applied to obtain a global semantic representation. It is used for querying.

[0110]

[0111] in .

[0112] Using cosine similarity Measuring the correlation between the input sequence and the soft instances stored in the instance store, using queries from Before searching The most relevant instance is represented as an explicit learning of relevant events to enrich the diversity of external instances.

[0113]

[0114] here These are the top K most relevant instances retrieved.

[0115] Adaptive fusion mechanism. In order to explicitly learn relevant events from the retrieved instance representations, this application designs an adaptive fusion mechanism to construct aggregated instance representations.

[0116] Specifically, this application converts the similarity score between the query and each retrieved instance representation into normalized attention weights. To generate memory output. Query. and each retrieved instance representation The cosine similarity S between them is calculated as follows:

[0117]

[0118] In the formula, Represents the relationship between the i-th vector and the query. Cosine similarity between them ;

[0119] The relevant attention weights are defined as:

[0120]

[0121] In the formula, Represents the relationship between the j-th vector and the query. Cosine similarity between them , It's a hyperparameter. These are the top K most relevant instances retrieved.

[0122] Finally, this application computes the aggregate memory representation. The weighted sum, representing the retrieved instances, is as follows:

[0123]

[0124] In the formula, .

[0125] 3. Construct the hybrid generator as follows:

[0126] To fully utilize both abstract conceptual information and concrete instance information, this application proposes a hybrid enhancement mechanism that improves the encoder representation by integrating retrieved instance information. This process incorporates instance-level contextual information while preserving the original semantic content. The enhanced encoder representation is computed as follows:

[0127]

[0128] In the formula, and Represents trainable parameters. .

[0129] In the generation step Multi-layer Transformer decoder predicts the first Output token and decoder hidden state as follows:

[0130]

[0131] in, Each layer is implemented as a Transformer decoder block, which includes the previously generated decoder state. Masking self-attention and representation of the augmented encoder Cross-attention. Therefore, the conditional probability of the entire output sequence is decomposed by autoregression as follows:

[0132]

[0133] here This indicates the number of tokens in the output sequence.

[0134] Training and reasoning

[0135] During training, this application uses document pairs and their corresponding event sequences to optimize the model. Given a training dataset... Each instance is composed of ( For the composition, the learning objective is to minimize the negative log-likelihood loss:

[0136]

[0137] During the reasoning process, given an input document HyGEN generates event sequences in an autoregressive manner, and then extracts the final event arguments and roles from these structured sequences through a template-based matching mechanism.

[0138] Furthermore, the above method was verified through experiments as follows: The experimental setup was as follows:

[0139] Table 1. Details of the dataset used in the experiment.

[0140]

[0141] The dataset is as follows:

[0142] This application conducts experiments on four standard event argument extraction datasets: RAMS, CMNEE, ACE 05-EN, and ACE 05-CH. RAMS is a document-level argument extraction dataset designed to evaluate a model's ability to retrieve and link event arguments across sentence boundaries. It originates from newsletter text and contains 9,124 annotated events covering 139 event types. CMNEE focuses on news text and includes approximately 17,000 documents containing 29,223 manually annotated event instances. It defines eight event types, providing a large-scale dedicated resource for document-level event extraction. ACE 05 is a manually annotated sentence-level benchmark dataset for general-domain event extraction. It is divided into eight coarse-grained event categories and 33 fine-grained subcategories. This application uses its English and Chinese subsets, referred to as ACE05-EN and ACE05-CH, respectively. More detailed information about these datasets is shown in Table 1.

[0143] The evaluation metrics and baselines are as follows:

[0144] This application follows previous work and uses Precision, Recall, and F1 as the primary evaluation metrics. Specifically, this application focuses on Arg-C argument classification: an argument is considered correct only if its text and role match the truth value. Let TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively. Precision, Recall, and F1 are calculated as follows: , , .

[0145] To ensure comparability with previous studies, this application reports the average scores for all argument roles.

[0146] To facilitate systematic comparisons, this application divides the baselines by dataset language. In the English dataset (RAMS, ACE05-EN), the following representative baselines were selected for experimental comparison:

[0147] CMR uses dynamic compressed memory to effectively cache retrieved information while enabling continuous memory updates.

[0148] FSGB integrates selective and generative approaches through fusion hints to form a unified modeling framework.

[0149] DEGAP enriches the input representation by using double-prefix interactive prompts (instance-oriented and prompt-oriented).

[0150] DEEIA is a multi-event argument extraction model that uses a multi-event cueing mechanism to obtain the interaction between events.

[0151] ERCL proposes a hierarchical error-aware contrastive learning framework that learns the distribution of potential extraction errors at the sentence and span levels.

[0152] TabEAE reformulates EAE as a table generation problem and extends it to a non-autoregressive generation framework for extracting multiple events in parallel.

[0153] AMPERE introduces abstract semantic representations to generate perceptual prefixes and introduces an adaptive adjustment mechanism.

[0154] APE decomposes EAE knowledge into cross-dataset overlapping knowledge and target-specific knowledge, and learns them sequentially through a two-stage conditional generation strategy.

[0155] In the Chinese dataset (CMNEE, ACE 05-CH), this application selects the following representative baselines:

[0156] HiEE combines hierarchical feature fusion and external semantic enhancement, and redefines the task as trigger word-argument relation classification.

[0157] MLAEE provides a shared label distribution for each Chinese character and incorporates a word dictionary into the character representation.

[0158] Dep-BERT embeds explicit syntactic constraints into an attention mechanism to guide event argument extraction.

[0159] DEGREE is designed to provide semantic guidance by summarizing the events described in an article into a natural sentence that follows a predefined template.

[0160] PAIE uses cue enhancement for argument extraction, introducing two cue-based span selectors to extract multiple arguments with the same role.

[0161] BERT+CRF uses an encoder to obtain a document-level context representation and finally applies a CRF layer to impose label transformation constraints on the label-level score and solve for the optimal label sequence.

[0162] EEQA introduces a new event extraction paradigm by reconstructing event argument extraction into an end-to-end question-answering task.

[0163] TEXT2EVENT proposes a sequence-to-structure generation paradigm that extracts events directly from text in an end-to-end manner.

[0164] The parameter details are as follows:

[0165] In all experiments, this application sets the learning rate to The batch size for training is set to 6, and the batch size for validation and testing is set to 12. Weight decay is fixed. All models were trained for 45 epochs. Each experiment was conducted using three random seeds (0, 23, and 42), and the reported results are the average of the experiments. This application uses BART-large for the English dataset and mBART-large for the Chinese dataset. All experiments were performed on an NVIDIA H100 GPU with 80 GB of memory.

[0166] The main experimental results are as follows:

[0167] Based on the main experimental results reported in Tables 2, 3 and 4, this application draws the following conclusions:

[0168] Results on English datasets. Table 2 shows the experimental results of this application on the RAMS and ACE 05-EN test sets. It can be observed that the proposed method, HyGEN, achieves state-of-the-art performance on various English datasets.

[0169] Table 2 Experimental results (F1%) of RAMS and ACE 05-EN

[0170]

[0171] In Table 2-4, the best results are highlighted in bold, and the second best results are indicated by underline.

[0172] On the document-level dataset RAMS, the proposed HyGEN improves the F1 score by 3.5% (54.3% → 57.8%) compared to previous state-of-the-art models. This result reflects HyGEN's ability to better distinguish semantically similar events while reducing cross-sentence argument omissions, thus achieving a more stable balance between accuracy and recall. In contrast, existing methods are more prone to distraction in complex contexts and struggle to fully utilize global semantic information. Meanwhile, on the sentence-level dataset ACE05-EN, the proposed HyGEN improves the F1 score by 0.8% (75.4% → 76.2%). Given the relatively simple sentence structure in this dataset and the already good performance of existing methods, this improvement demonstrates that HyGEN can still effectively enhance sentence-level argument extraction.

[0173] Furthermore, HyGEN from this application outperforms all RAG methods, exceeding the best RAG method CMR's F1 score by 3.5% and 1.4% on RAMS and ACE05-EN, respectively. Simultaneously, compared to existing generative methods, HyGEN surpasses the best model APE's F1 score by 3.5% and 0.8% on RAMS and ACE05-EN, respectively. Overall, HyGEN demonstrates its core competence in argument extraction through cross-framework comparisons.

[0174] Table 3 Experimental results of CMNEE (F1%)

[0175]

[0176] Table 4 Experimental results of ACE05-CN (F1%)

[0177]

[0178] Results on Chinese datasets. Tables 3 and 4 show the experimental results of this application on the CMNEE and ACE 05-CH test sets. It can be observed that the method HyGEN in this application achieves more significant performance improvements on different Chinese datasets.

[0179] On the document-level dataset CMNEE, HyGEN improved the F1 score by 5.2% (69.4% → 74.6%) for frequently occurring complex names and long, difficult sentence structures. This indicates that HyGEN has strong entity boundary recognition capabilities when handling low-frequency technical terms, effectively overcoming the problems posed by long and complex sentences in news articles. On the ACE 05-CH dataset, which contains corpora from multiple sources including news, broadcasts, and online comments, HyGEN achieved a consistent 5.2% (60.4% → 65.6%) improvement in F1 score, effectively validating its adaptability across different scenarios.

[0180] Overall Performance. Experimental results on four public datasets (ACE 05-EN, RAMS, ACE 05-CH, CMNEE) demonstrate that our proposed HyGEN achieves state-of-the-art performance. Compared to various existing models, our HyGEN exhibits significant improvements, with a 3.5% improvement in F1 score on RAMS, a 0.8% improvement on ACE 05 EN, and a 5.2% improvement on CMNEE and ACE 05-CH. Our experimental results also show that our HyGEN achieves more significant performance gains in Chinese than in English. Furthermore, the fact that document-level improvements outperform sentence-level improvements underscores our HyGEN's ability to understand complex contexts.

[0181] In summary, the experimental results highlight HyGEN's versatility and effectiveness across multiple dimensions. First, the improved consistency between English and Chinese demonstrates its cross-linguistic compatibility, indicating that HyGEN effectively captures latent semantic features. Furthermore, significant improvements at both the document and sentence levels (especially the latter) validate its cross-granularity compatibility.

[0182] Table 5 Ablation experiments (%) on RAMS, ACE05-EN and ACE05-CH

[0183]

[0184] The ablation experiment is as follows:

[0185] To evaluate the contribution of each proposed component to the overall performance, this application conducted ablation experiments on the test sets of the RAMS, ACE05-EN, and ACE05CH datasets. This application compared three settings: (1) w / o FCGP, which excludes fine-grained concept guidance hints and relies solely on documents and event templates as input; (2) w / o SIAM, which omits soft instance-aware memory, resulting in direct encoding to decoding without memory retrieval; and (3) w / o All, which removes both fine-grained concept guidance hints and soft instance-aware memory, leaving an encoder-decoder architecture that only uses documents and templates as input. The results of the ablation experiments are shown in Table 5. Where w / o FCGP = no fine-grained concept guidance hints, w / o SIAM = no soft instance-aware memory, and w / o All = no fine-grained concept guidance hints and soft instance-aware memory.

[0186] Based on the ablation results, this application conducts a detailed analysis to explore the contributions of different components. (1) Soft instance-aware memory plays a crucial role. Removing soft instance-aware memory leads to a significant decrease in F1 scores across all three datasets: from 57.76% to 56.53% (a decrease of 1.23%) on RAMS, from 76.20% to 74.10% (a decrease of 2.10%) on ACE05-EN, and from 65.58% to 62.21% (a decrease of 3.37%) on ACE05-CH. These results indicate that soft instance-aware memory has universal effectiveness across different domain data and languages. The continued performance decline after removal confirms that this component captures underutilized instance-level information. (2) Fine-grained concept guidance cues exhibit significant data dependence. Fine-grained concept guidance resulted in a 0.6% F1 improvement on RAMS, a 0.3% F1 improvement on ACE05-EN, and zero gain on ACE05-CH (remaining at 65.6%). The fundamental reason for this difference lies in the varying complexity of the datasets. Specifically, RAMS contains complex documents where fine-grained guidance is crucial for helping the model distinguish difficult arguments. On ACE05-EN, arguments are relatively easier to extract, leading to smaller improvements because the model's existing functionality has already been largely extracted. For the simpler dataset ACE05-CH, event arguments can already be handled effectively without fine-grained concept guidance, making the additional guidance redundant. These results indicate that this component is primarily effective for challenging instances, and its effectiveness gradually decreases as the context becomes simpler.

[0187] Experimental results demonstrate that the HyGEN model with all components achieves state-of-the-art performance across all datasets. This result suggests that the performance improvement observed in our experiments is largely attributable to two key components: fine-grained concept guidance cues and soft instance-aware memory. The former explicitly injects semantic information into the model input, providing accurate prior guidance for argument extraction. The latter utilizes instance-level memory to enhance representations and suppress noise from similar but irrelevant events. The synergistic effect of these two components significantly enhances the model's capabilities in DoCEAE.

[0188] The impact of the number of searches, k, is as follows:

[0189] Table 6 reports instances where the retrieval was changed. The performance changes across datasets depending on the size of the dataset. The results show that the amount of memory retrieved has a significant impact on model performance.

[0190] Specifically, on RAMS, the model is... At a resolution of 6, the model achieved a peak F1 score of 57.76%. As a document-level dataset, RAMS requires a large retrieval scale to capture sufficient instance-level information. In contrast, on ACE05-EN, the model achieved a much smaller retrieval scale. = 4 achieved a best F1 score of 76.20%, indicating that the retrieved instance size is sufficient to provide the necessary contextual information for this benchmark. For ACE05-CH, performance is at The performance peaks at k=2, with an F1 score of 65.58%, and increasing k leads to a performance decrease. This suggests that in simple sentence-level datasets where contextual information is already sufficient, fewer but highly relevant retrieval instances are more beneficial.

[0191] Table 6 When the retrieved instances are changed Results (%) on RAMS, ACE05-EN, and ACE05-CH at different sizes.

[0192]

[0193] In general, the size of the retrieved memory instances plays a crucial role in balancing representation enhancement and noise introduction. A moderate size, typically between 2 and 6, is preferred. This often achieves optimal performance, allowing HyGEN to effectively utilize external memory without overfitting irrelevant instances. Different datasets exhibit different optimal performance due to variations in event complexity. Simple sentence-level datasets (such as ACE05-CH) require fewer instances, while more complex datasets (such as ACE05-EN and document-level RAMS) benefit from larger... This allows for a better acquisition of comprehensive semantic cues. These findings suggest that... It should be adaptively adjusted according to the complexity of the text, providing a practical heuristic for optimizing model performance across different scenarios.

[0194] The comparative experiment with LLMs is as follows:

[0195] To demonstrate the superiority of the proposed method, this application compares its model with several publicly available LLMs, including Qwen3-30b, Qwen3-max, and DeepSeek-V3.2Exp. All comparisons follow the same data segmentation and evaluation metrics (precision / recall / F1) used in HyGEN. ACE05EN and ACE05-CH both use the same test data to ensure experimental comparability. To further guarantee experimental comparability and setup consistency, all LLMs use the same prompts as the fine-grained conceptual guidance prompts proposed in HyGEN. Furthermore, a retrieval-enhanced few-shot mechanism is employed. Here, the few-shot count is set to 5. Unlike traditional static few-shot settings, few-shot instances are dynamically retrieved from the training corpus based on the consistency between the retrieved event type and the current input event type. These retrieved instances are integrated into the context, forming a retrieval-enhanced few-shot, which is conceptually consistent with the soft instance-aware memory proposed in HyGEN. It is worth noting that, given the dynamic nature of online APIs, all LLM experimental results reported in this study represent the performance of the models as of November 14, 2025.

[0196] Table 7 shows the results (%) of comparing the HyGEN proposed in this application with LLMs on the ACE05-EN and ACE05-CH datasets.

[0197]

[0198] All LLMs were evaluated in 5-few-shots.

[0199] Table 7 summarizes the performance comparison of the proposed HyGEN with several LLMs on the ACE05-EN and ACE05-CH datasets. The results show that although Qwen3-max achieves the highest performance on both datasets, its overall performance is still lower than the model proposed in this application. On the ACE05-EN dataset, the model proposed in this application achieves Precision = 73.00%, Recall = 79.70%, and F1 = 76.20%, which means an absolute improvement of +27.21% in F1 score compared to Qwen3-max, and significant improvements of approximately +27.47% and +26.63% in Precision and Recall scores, respectively. Compared to Qwen3-max, the model proposed in this application achieves Precision = 64.39%, Recall = 66.81%, and F1 = 65.58%, and on the ACE05-CH dataset, the absolute improvement in F1 score is +21.81%. Furthermore, compared to Qwen3-max, the model in this application improves precision and recall by approximately 19.77% and 23.86%, respectively, demonstrating a significant advantage in argument extraction accuracy and recall.

[0200] Overall, compared with several state-of-the-art LLMs, the model in this application achieves significant performance advantages on both benchmark datasets, demonstrating superior effectiveness and broad applicability of argument extraction.

[0201] This application also provides a hybrid prompt generation apparatus for document-level event argument extraction, comprising:

[0202] The first unit is used to design fine-grained concept guidance prompts, which include event types, event trigger words and location information, as well as argument roles;

[0203] The second unit is used to construct soft instance-aware memory to store representations of all training instances and selectively retrieve information. Soft instance-aware memory includes memory construction, memory retrieval, and adaptive fusion mechanisms.

[0204] The third unit is used to build a hybrid generator based on the decoder. The hybrid generator is used to generate arguments by fusing fine-grained concept guidance prompts and soft instance-aware memory.

[0205] The fourth unit is used to build a hybrid prompt generation model based on fine-grained concept guidance prompts, soft instance-aware memory, and hybrid generators;

[0206] The fifth unit is used for argument extraction based on a hybrid prompting generation model and real-time documents.

[0207] The hybrid prompt generation apparatus for document-level event argument extraction described above can implement various embodiments of the hybrid prompt generation method for document-level event argument extraction described above, and can achieve the same beneficial effects. Here, it will not be elaborated further.

[0208] This application also provides a hybrid prompt generation apparatus for document-level event argument extraction, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method. The above-described hybrid prompt generation apparatus for document-level event argument extraction can implement various embodiments of the above-described hybrid prompt generation method for document-level event argument extraction and achieve the same beneficial effects; therefore, further details are omitted here.

[0209] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0210] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hybrid prompt generation method for document-level event argument extraction, characterized in that, The method includes: Design fine-grained concept guidance prompts, which include event type, event trigger words and location information, as well as argument roles; A soft instance-aware memory is constructed to store representations of all training instances and selectively retrieve information. The soft instance-aware memory includes memory construction, memory retrieval, and an adaptive fusion mechanism. A hybrid generator is built based on the decoder, which is used to generate arguments by fusing fine-grained concept guidance prompts and soft instance-aware memory; A hybrid prompt generation model is constructed based on fine-grained concept guidance prompts, soft instance-aware memory, and a hybrid generator; Argument extraction is performed based on the hybrid prompt generation model and real-time documents; The construction process of the memory structure is as follows: Given a training corpus ,Depend on It consists of documents, each document... The input format is converted to a unified form as follows: In the formula, , These represent the start and end characters, respectively. The input is fed into the shared encoder to obtain the start character. Context representation It is used as an instance representation. ,as follows: in All instances represent the memory that collectively forms a soft instance. ,as follows: ； The construction process of the memory retrieval is as follows: Instance-based memory and the context representation of the input ,exist Max pooling is applied to obtain a global semantic representation. It is used for querying, as follows: in ; Using cosine similarity Measuring the correlation between the input sequence and the soft instances stored in the instance store, using queries from Before searching The most relevant instances are represented as explicit learning of relevant events, as follows: In the formula, These are the top K most relevant instances retrieved; The adaptive fusion mechanism is used to construct aggregated instance representations, and the construction process is as follows: Query The similarity score between each retrieved instance representation is converted into a normalized attention weight. To generate memory output, where the query and each retrieved instance representation The cosine similarity S between them is calculated as follows: In the formula, Represents the relationship between the i-th vector and the query. Cosine similarity between them ; The relevant attention weights are defined as: In the formula, Represents the relationship between the j-th vector and the query. Cosine similarity between them , It's a hyperparameter. These are the top K most relevant instances retrieved; Computational aggregate memory representation The weighted sum, representing the retrieved instances, is as follows: In the formula, .

2. The hybrid cue generation method for document-level event argument extraction according to claim 1, characterized in that, The event type is used to provide a fine-grained semantic description of the event type; the event trigger word and location information are used to specify the trigger word and its corresponding position in the input document; The argument roles are used to provide natural language templates with role placeholders.

3. The hybrid prompt generation method for document-level event argument extraction according to claim 1, characterized in that, The fine-grained concept guidance prompts satisfy the following relationship: in, This indicates a splicing operation. It is an event type guideline. It is a trigger word guide prompt. It is an argument role guide prompt; The method further includes: Input document With corresponding fine-grained concept guidance tips Connect to form the encoder input, as follows: The contextual representation of the input is computed using a multi-layer Transformer encoder. ,as follows: in , Indicates the maximum sequence length of the encoder. Indicates the hidden dimension. Each layer is implemented as a Transformer encoder block with a self-attention mechanism.

4. The hybrid prompt generation method for document-level event argument extraction according to claim 3, characterized in that, The decoder-based hybrid generator construction includes: The computationally enhanced encoder is represented as: In the formula, and Represents trainable parameters. ; In the generation step Multi-layer Transformer decoder predicts the first Output token and decoder hidden state as follows: in, Each layer is implemented as a Transformer decoder block, which includes the previously generated decoder state. Masking self-attention and representation of the augmented encoder Cross attention; The conditional probability of the entire output sequence is decomposed by autoregression as follows: In the formula, This indicates the number of terms in the output sequence.

5. The hybrid cue generation method for document-level event argument extraction according to claim 4, characterized in that, The learning objective of the hybrid prompt generation model is to minimize the negative log-likelihood loss: In the formula, Let be the conditional probability of the entire output sequence. This is the training dataset.

6. A hybrid prompt generation apparatus for document-level event argument extraction, characterized in that, include: The first unit is used to design fine-grained concept guidance prompts, which include event types, event trigger words and location information, as well as argument roles; The second unit is used to construct a soft instance-aware memory for storing representations of all training instances and selectively retrieving information. The soft instance-aware memory includes memory construction, memory retrieval, and an adaptive fusion mechanism. The third unit is to construct a hybrid generator based on the decoder. The hybrid generator is used to generate arguments by fusing fine-grained concept guidance prompts and soft instance-aware memory. The fourth unit is used to build a hybrid prompt generation model based on fine-grained concept guidance prompts, soft instance-aware memory, and hybrid generators; The fifth unit is used for argument extraction based on the hybrid prompt generation model and real-time documents; The construction process of the memory structure is as follows: Given a training corpus ,Depend on It consists of several documents, each document... The input format is converted to a unified form as follows: In the formula, , These represent the start and end characters, respectively. The input is fed into the shared encoder to obtain the start character. Context representation It is used as an instance representation. ,as follows: in All instances represent the memory that collectively forms a soft instance. ,as follows: ； The construction process of the memory retrieval is as follows: Instance-based memory and the context representation of the input ,exist Max pooling is applied to obtain a global semantic representation. It is used for querying, as follows: in ; Using cosine similarity Measuring the correlation between the input sequence and the soft instances stored in the instance store, using queries from Before searching The most relevant instances are represented as explicit learning of relevant events, as follows: In the formula, These are the top K most relevant instances retrieved; The adaptive fusion mechanism is used to construct aggregated instance representations, and the construction process is as follows: Query The similarity score between each retrieved instance representation is converted into a normalized attention weight. To generate memory output, where the query and each retrieved instance representation The cosine similarity S between them is calculated as follows: In the formula, Represents the relationship between the i-th vector and the query. Cosine similarity between them ; The relevant attention weights are defined as: In the formula, Represents the relationship between the j-th vector and the query. Cosine similarity between them , It's a hyperparameter. These are the top K most relevant instances retrieved; Computational aggregate memory representation The weighted sum, representing the retrieved instances, is as follows: In the formula, .

7. A hybrid cue generation apparatus for document-level event argument extraction, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any of the methods described in claims 1-5.

Images