An event argument recommendation method based on multi-model fusion, an electronic device, and a storage medium

Abstract

The application belongs to the technical field of information extraction, and provides an event argument recommendation method based on multi-model fusion, an electronic device and a storage medium. The method comprises the following steps: inputting a text, an event trigger word, an event type, an arbitrary event-defined role set, and each argument role; inputting each argument role in the role set into a plurality of argument extraction sub-models of different technical routes to obtain a plurality of predicted arguments and a plurality of candidate arguments corresponding to the confidence of the plurality of candidate arguments; forming a candidate argument set by using the plurality of predicted arguments and the confidence corresponding to the plurality of candidate arguments; and outputting a recommended argument set containing a plurality of results for any argument role in the candidate argument set. The method generates candidates in parallel through a plurality of heterogeneous models, and then obtains the final recommended result through confidence alignment, calculation of fusion scores and diversity sorting, so that the correct argument appears in the Top-K candidate with a high probability for quick confirmation by a human.

Description

Technical Field

[0001] This invention belongs to the field of information extraction technology, and relates to an event argument recommendation method based on multi-model fusion, an electronic device, and a storage medium. Background Technology

[0002] With the rapid development of information technology, a large amount of text data is constantly emerging. How to accurately extract valuable event information from this text has become a research hotspot in the field of natural language processing. Traditional event argument extraction models aim to achieve fully automated extraction tasks, outputting the most likely event arguments for each event.

[0003] In practical applications, such as legal document extraction and disaster event monitoring, although traditional models can extract event arguments to a certain extent, they have many shortcomings: ① Due to the complexity and diversity of natural language, the results of automatic extraction by traditional models often contain errors. Problems such as polysemy and semantic ambiguity in text make it difficult for models to accurately understand the meaning of the text, resulting in inaccurate extraction of event arguments or missed detections.

[0004] ② Traditional models lack effective collaboration with human intervention. In many scenarios where the accuracy of extraction results is extremely critical, human review and correction are essential. However, since traditional models are primarily designed for automated extraction with a single output, if a human detects an error, they need to reread the entire text to find the correct event arguments. This significantly increases the workload and time cost for human intervention. For example, in event extraction from legal documents, the accuracy of event elements is extremely important; if the model extracts incorrect information, manual review is time-consuming and laborious.

[0005] In summary, traditional event argument extraction models face numerous challenges in practical applications and cannot meet the demands for efficient and accurate human-machine collaborative event argument extraction. Therefore, constructing a human-machine collaborative event argument recommendation model, innovatively designing model reasoning and computation mechanisms, recommending multiple candidate results to ensure high recall, and providing options for manual review and selection are of significant practical importance for improving the efficiency and quality of human-machine collaboration.

[0006] Existing event argument extraction methods are primarily designed for automated extraction tasks, typically outputting a single result for each event argument role, resulting in limited recall. In tasks requiring high-quality results (such as legal document analysis and emergency event monitoring), human review of the model's extraction results is necessary. However, due to insufficient model recall, manual review often requires rereading the entire text to find missing arguments, leading to a heavy workload for humans and low efficiency in human-machine collaboration. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention aims to design an event argument recommendation method based on multi-model fusion, which recommends multiple candidate results for each event argument role, significantly improving recall rate, allowing for rapid selection and verification by humans, reducing the need to repeatedly read the full text, and thus improving the overall efficiency of human-machine collaborative event argument extraction.

[0008] This invention provides an event argument recommendation method based on multi-model fusion, comprising the following steps: Step 1: Let the input text be... Event trigger words Event Type The set of roles defined for any type of event is and each argument role ; Step 2: Assemble the characters Each argument role in The input is fed into multiple argument extraction sub-models with different technical approaches, and multiple predicted arguments and their corresponding confidence scores are obtained respectively. Step 3: Combine multiple predicted arguments and their corresponding confidence scores to form a candidate argument set. ; Step 4: Set of candidate arguments Any argument role in The output includes The set of recommendation arguments for each result.

[0009] Furthermore, several sub-models for argument extraction using different technical approaches include question-answering guided extraction models, template-filling extraction models, and dialogue-generating models.

[0010] Furthermore, the specific process of generating predicted arguments using the question-answer guided extraction model is as follows: S2.11 for each argument role Constructing its natural language problem ; S2.12, Natural Language Problems With input text The data is concatenated together and used as input to the pre-trained model in the question-answering guided extraction model. ; S2.13, based on Obtain the context representation ; S2.14, Context-based representation Construct the head pointer Sum of tail pointers And based on the head pointer Sum of tail pointers Determine any argument role starting position and end position ; S2.15, Set the candidate text fragments of the event arguments as follows: Based on any argument role starting position and end position Calculate the set of candidate text fragments Combined score ; S2.16, Joint score based on candidate text fragments Select the highest-scoring legitimate text segment as the model's prediction argument. ; S2.17 Predicted arguments for all candidate text segments Perform Softmax normalization to obtain its confidence score. .

[0011] Furthermore, the specific process of generating predictive arguments using the template-filling extraction model is as follows: S2.21 Input Construction; The construction includes fill-in-the-blank natural language prompt templates for all roles. ; S2.22. Using a model with an Encoder-Decoder architecture, the input text is processed... This serves as the input and prompt template for the Encoder part of the model. As input to the Decoder part of this model, an event-oriented context representation is obtained. and context-oriented hints ; S2.23, Define the role of each argument. have One slot; S2.24 Context-Oriented Hint Representation Get each argument role The Role characteristics of each slot ;in, ; S2.25, From the perspective of character traits The role of each argument is derived from the derivation. The Text snippet selector for each slot ; Use text fragment selector From the context of event-oriented representation Extract a corresponding event argument; S2.26. Repeat steps S2.24 and S2.25 to obtain the role of each argument. The event arguments corresponding to all slots are used to define the resulting event arguments as a set of candidate text fragments. ; For the candidate text fragment set Score for each text segment Perform calculations; S2.27, Scoring based on each text segment The text segment with the highest score was selected as the prediction argument. ; S2.28 Predicted arguments for all candidate text segments Perform Softmax normalization to obtain the predicted arguments. confidence level .

[0012] Furthermore, the specific process of generating predicted arguments using a modal dialogue generative model is as follows: S2.31, Constructor Instruction Hints ; S2.32, Define the role of each argument. The generated prediction arguments are ; Based on the intermediate information generated during decoding, the generation probability is... ; S2.33, Based on the generation probability Calculate the role of each argument Includes the average log probability of each word and assigns each argument role The average log probability of the included word as the generation confidence. .

[0013] Furthermore, the specific method of step three is as follows: For each argument role Its candidate argument set ,in, For the corresponding text fragment, For model confidence, As the source of the model, These are the aligned candidate arguments.

[0014] Furthermore, the specific process of step four is as follows: S4.1. Let any one of the argument extraction sub-models from multiple different technical approaches be... The extreme value standardization method is used for any sub-model. Normalization is performed to obtain the normalized confidence level. ; S4.2, Based on normalized confidence level Calculate each aligned candidate argument Fusion score ; S4.3 From the set of candidate arguments The Top-K recommended items are selected, and the maximum marginal relevance algorithm is used to balance relevance and diversity to obtain the argument roles. of Recommended results .

[0015] Furthermore, we obtain the argument role. of Recommended results The specific process is as follows: Recommendation results Perform initialization: ; in, It is an empty set. This is the set of candidate arguments that were not recommended. Iteration to ; ; ; in, This is a function for calculating the similarity between two candidate arguments, obtained by calculating the cosine similarity of the semantic vectors corresponding to the candidate arguments; , The semantic vector of the candidate arguments after being encoded by the BERT pre-trained model; For the k-th recommended candidate argument, For argument confidence weights in recommendation algorithms, As candidate arguments, As another candidate argument; Based on the argument role of Recommended results ; .

[0016] The present invention also provides an electronic device including a memory, one or more processes, and one or more programs stored in the memory, said one or more programs including instructions for performing the event argument recommendation method as described above.

[0017] The present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, said one or more programs including instructions for performing the event argument recommendation method as described above.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides an event argument recommendation method based on multi-model fusion. First, given text... Event trigger words Event Type and the set of argument roles for this type of event. The data is input into multiple argument extraction sub-models using different technical approaches, which extract corresponding candidate arguments. The confidence of each candidate argument is calculated using a confidence calculation module, and these are combined to form a candidate argument set. Then, through multi-model confidence alignment, candidate argument score fusion calculation, and a confidence-diversity-aware TopK recommendation algorithm, each argument is ultimately assigned a role. The output includes This method generates a set of recommended arguments with high recall and low redundancy. Candidates are generated in parallel using multiple heterogeneous models, and then the final recommendation results are obtained through confidence alignment, calculation of fusion scores, and diversity ranking. This ensures that correct arguments appear with high probability in the Top-K candidates for rapid human confirmation.

[0019] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0020] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating an event argument recommendation method based on multi-model fusion in Embodiment 1 of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention clearer and easier to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in a simplified form and use non-precise proportions, and are only used to facilitate and clearly illustrate the implementation of the present invention.

[0022] Example 1: To improve the recall rate of event arguments in text, this invention provides an event argument recommendation method based on multi-model fusion, see [link to relevant documentation]. Figure 1As shown, it includes the following steps: Step 1: Let the input text be... Event trigger words Event Type And the set of roles defined for this type of event is On meta-roles The goal of each model is to output an argument. And give the corresponding confidence level. To form a tuple .

[0023] Preferably, input text The expression is as follows: ; in, This represents the total number of texts.

[0024] Preferred event trigger words The expression is as follows: .

[0025] Preferably, the set of roles defined for this type of event The expression is as follows: ; in, This represents the total number of characters.

[0026] For example, on the evening of June 17, 2024, local time, in the fifth game of the basketball tournament finals held in a certain city, Team A defeated Team B 106-88 and won the championship with a series score of 4-1. Player M1 scored 31 points, 11 assists and 8 rebounds, while player M2 contributed 21 points and was ultimately selected as the Finals MVP.

[0027] Trigger word: defeat; Time type: Sports competition - win / loss; Role set: {Time; Winner; Loser; Score}.

[0028] Step 2: Assemble the characters Each argument role in The input is fed into multiple argument extraction sub-models with different technical approaches, and multiple predicted arguments and their corresponding confidence scores are obtained.

[0029] As a further solution in this embodiment, since single models are often limited by their inherent technical bottlenecks and data biases in the event argument extraction task, it is difficult to comprehensively and robustly capture the complex semantic relationships in the text. To overcome this limitation, this embodiment adopts a multi-perspective and multi-technical approach fusion strategy, integrating three models with differences and complementarities. The multiple argument extraction sub-models with different technical approaches include Question-Guided Extraction (QGE, i.e., Model 1), Template-based Infilling Extraction (TIE, i.e., Model 2), and Prompt-based Generative Model (PGM, i.e., Model 3). The three complement each other in terms of precise localization, structural constraints, and semantic reasoning: QGE and TIE, as discriminative models, provide high-confidence explicit argument candidates, ensuring the accuracy of the results; PGM, as a generative model, supplements implicit arguments and unconventional expressions, significantly improving recall; at the same time, QGE's question-answering mechanism and TIE's template constraints mutually verify each other, filtering out possible hallucinatory outputs from PGM, while PGM's semantic understanding can make up for QGE / TIE's shortcomings in handling fuzzy references.

[0030] The preferred method for constructing a question-guided extraction model is as follows: For the question-answering guided extraction model, a pre-trained language model (BERT is preferred in this embodiment) is used as a shared encoder to perform deep semantic encoding on the input text sequence; Construct two isomorphic but parameter-independent linear projection layers, which serve as the start position predictor and the end position predictor, respectively.

[0031] For example, Question 1: When was the match held? Question 2: Who won? Question 3: Who was defeated?

[0032] The preferred method for constructing a template-filling extraction model is as follows: A pre-trained model with an Encoder-Decoder architecture (preferably BART in this embodiment) is used as the base model. This model is responsible for encoding the input text into hidden features and decoding it to generate the target text. For a predefined set of event types, a prompt template is constructed for each event type. For example, for the "sports competition - win / loss" event, the template is designed as: "At [match time], [the winning team] defeated [the losing team] by [score]". The placeholders in the template correspond one-to-one with the specific argument roles. A pair of text fragment selector parameters shared by the roles are introduced. This selector is a pair of learnable embedding vectors with the same dimension as the model's word vectors.

[0033] Preferably, the construction process of the dialogue generative model is as follows: A generative large language model is selected (in this embodiment, the Qwen series of open-source models is preferred). This model is responsible for understanding user commands and generating natural language responses containing argument information. An event-type adaptive prompt word template library is constructed, which contains static system commands and dynamic event descriptions. The static commands are used to define the model's role (e.g., "You are an event extraction expert"); the dynamic part automatically inserts the corresponding structured task description based on the specific event type of the input text, for example: "Please extract the match result event with 'defeat' as the trigger word from the text, and output [match time], [winner], [loser], and [score]." As a further embodiment, the specific process of generating initial candidate arguments (i.e., predicted arguments) using a question-guided extraction (QGE) model is as follows: S2.11 for each argument role Constructing Natural Language Problems ,in, This serves as a question template. For example, for the "winner" argument in the "sports competition - win or lose" event, the question is "In the [trigger word]-represented [sports competition - win or lose] event, who won?".

[0034] S2.12, Natural Language Problems With input text The data is concatenated together and used as input for a question-guided extraction model. ; Preferred, The expression is as follows: ; in, This is a placeholder in a pre-trained model, usually placed at the beginning and can be used for classification tasks. These are placeholders in the pre-trained model, used to join the question and text.

[0035] S2.13, will Input a pre-trained model (specifically, in this embodiment, the pre-trained model is set to BERT) to obtain a context representation. ; Context representation The expression is as follows: ; in, For the set of real numbers, For context length, For contextual dimensions.

[0036] S2.14, In context Based on this, a pair of head and tail pointers (two learnable linear projection heads) are constructed to determine the start and end positions of arguments; Preferred, head pointer The expression is as follows: ; ; ; in, This is the bias parameter for the head pointer. The weight matrix of the head pointer, with superscript... This indicates the matrix transpose.

[0037] Preferably, the tail pointer The expression is as follows: ; ; ; in, This is the offset parameter for the tail pointer. The weight matrix is ​​for the tail pointer.

[0038] S2.15, Set the candidate text fragments of the event arguments as follows: Based on any argument role starting position and end position Calculate the set of candidate text fragments Combined score ; Preferred candidate text fragment set The expression is as follows: ; in, As the starting index, The text segment that ends the index.

[0039] Preferred, combined score The expression is as follows: ; in, for The first in The starting position of each element represents the token; for The first in The starting position of each element represents the token.

[0040] S2.16, Joint score based on candidate text fragments Select the highest-scoring legitimate text segment as the model's prediction argument. ; Model predictive arguments The expression is as follows: .

[0041] S2.17, Confidence Calculation: To generate confidence information for subsequent fusion score calculation and argument recommendation based on confidence, predicted arguments for all candidate text segments are generated. Softmax normalization is performed to obtain the predicted arguments of the model. The corresponding confidence level .

[0042] Preferred confidence level The expression is as follows: .

[0043] As a further embodiment, the specific process of generating initial candidate arguments (i.e., predicted arguments) using a template-based infilling extraction (TIE) model is as follows: S2.21 Input Construction; The construction includes fill-in-the-blank natural language prompt templates for all roles. ; Preferably, the expression for the construction is as follows: ; Specifically, the core idea of ​​the construction is to treat each slot, such as "[Winner]", as a learnable query parameter.

[0044] S2.22. Using a model with an Encoder-Decoder architecture, the input text is processed... Input and prompt templates for the Encoder section As input to the Decoder part, an event-oriented context representation is obtained. and context-oriented hints Specifically, in this embodiment, the BART model is preferred for models with an Encoder-Decoder architecture, meaning the input text... Enter BART-Encoder respectively, and the prompt template will appear. Enter BART-Decoder respectively, and the following prompt template will appear. and input text They will interact at the cross-attention layer of the decoder module.

[0045] Preferred, event-oriented context representation The expression is as follows: ; ; Preferred, context-oriented hint representation The expression is as follows: ; in: This refers to the encoder part of the BART model. This is a representation of the BART encoder model. This is the decoder part of the BART model.

[0046] S2.23, Regarding the first in the compound prompt words Each slot (corresponding to each argument role) ); S2.24, From the perspective of context-oriented prompt representation The pooling of the representation corresponding to each slot yields the feature of each character. ,in, Let be the dimension of the hidden layer in the BART model. .

[0047] Note that a character may have multiple slots in a given template, as well as multiple character traits and text snippet selectors.

[0048] S2.25, From the characteristics of each character The generation of each argument role is derived from the derivation. The Text snippet selector for each slot .

[0049] Preferred, each argument role The Text snippet selector for each slot The expression is as follows: ; ; ; ; in, Learnable parameters shared by all characters. For element-wise multiplication, For the prompt in the first A text snippet selector for each slot. For the first Start position selector for each slot A selector for the start position of a shared text fragment. For the first End position selector for each slot, A selector for the end position of a shared text segment.

[0050] This embodiment has only one shared learnable parameter. Combined with character traits It can generate different text fragment selectors and extract relevant arguments from the context, which allows the interaction between different roles and the interaction between the context and the roles to be considered in this paradigm.

[0051] Adopting the first Text snippet selector for each slot From the context of event-oriented representation Extract a corresponding event argument, namely the first one. The starting position of each slot in the context record. and end position distributed; Preferred starting position The expression is as follows: ; End position The expression is as follows: .

[0052] S2.26, Define the set of candidate text fragments for event arguments. ; For the candidate text fragment set Score for each text segment Perform calculations; ; S2.27, Scoring based on each text segment The text segment with the highest score was selected as the prediction argument. ; Predictive Argument The expression is as follows: ; S2.28 Predicted arguments for all candidate text segments Perform Softmax normalization to obtain the predicted arguments. confidence level ; Confidence The expression is as follows: .

[0053] As a further embodiment, the method of generating initial candidate arguments (i.e., predicted arguments) using a modal dialogue generative model (Prompt-based Generative Model, PGM) is as follows: Leveraging the language understanding and generation capabilities of generative models (such as open-source models like qwen), arguments are predicted in the form of dialogues.

[0054] A further preferred approach is to use a prompt-based generative model (PGM) to generate initial candidate arguments, as follows: S2.31, Constructor Instruction Hints ; Command prompts The expression is as follows: "[Original text]. Please extract the match win / loss events triggered by 'defeat' from the text, and output [match time], [winning side], [losing side], and [score]." S2.32, Define the role of each argument. The generated prediction arguments are To obtain the intermediate information of the model during decoding, the token generation probability is... ;in, To predict argument length, These are the prompt words input to the generative model; S2.33, Based on the token generation probability Calculate the role of each argument Includes the average log probability of each word and assigns each argument role The average log probability of the included word as the generation confidence. ; Preferred confidence level The expression is as follows: .

[0055] Step 3: Combine the predicted arguments and their confidence scores generated by the question-answering guided extraction model, the template-filling extraction model, and the dialogue-generating model to form a candidate argument set. Specifically, for each argument role Its candidate argument set ,in, For the corresponding text fragment, For model confidence, As the source of the model, These are the aligned candidate arguments.

[0056] For example: Time: (Evening of June 17, 2024, 1.9, Model 1); (Evening of June 17, 1.1, Model 2); (Fifth game of the finals, 0.5, Model 3); Winning teams: (Team A, 1.72, Model 1); (Team B, 0.32, Model 2); (Overall champion, 0.15, Model 3). The losing side: (Team A, 0.25, Model 1); (Team B, 1.65, Model 2); Score: (106 to 88, 0.95, Model 1); (4 to 1, 0.55, Model 2); (31 points, 0.15, Model 3).

[0057] Step 4: After multi-model confidence alignment, candidate argument score fusion calculation, and a confidence-diversity-aware TopK recommendation algorithm, the final candidate argument set is obtained. Each character in The output includes A set of recommendation arguments with high recall and low redundancy for each result.

[0058] Preferably, this section aims to fuse and rank initial candidate arguments from different models based on confidence levels, generating a high-quality, high-coverage unified candidate list. Specifically, for each role... Specifically, it includes the following sub-steps: S4.1 Confidence normalization; Since the confidence distributions of different models vary, they need to be normalized to a uniform scale. Therefore, in this embodiment, any one of the multiple argument extraction sub-models using different technical approaches is assumed to be... The extreme value standardization method is used for any sub-model. Normalization is performed to obtain the normalized confidence level. ;in, .

[0059] Preferred, normalized confidence level The expression is as follows: ; in, This represents the maximum confidence score of the model across all outputs in the current text. This represents the minimum confidence level of the model across all outputs in the current text.

[0060] S4.2, Calculation of integrated scores and conflict resolution; For each candidate argument after confidence normalization Calculate its fusion score .

[0061] Preferred, fusion score The expression is as follows: ; in, To support candidate arguments A collection of models; For the model In the role of the argument The scoring weights on the surface This is a custom parameter; the default value is the average.

[0062] S4.3, Confidence-based Top-K recommendation based on diversity perception; Finally from the set of candidate arguments The top-K recommended items are selected, and the maximum marginal relevance (MMR) algorithm is used to balance relevance and diversity.

[0063] The specific process is as follows: initialization: ; in, It is an empty set. This is the set of candidate arguments that were not recommended. Iteration to ; ; ; in, This is a function for calculating the similarity between two candidate arguments, obtained by calculating the cosine similarity of the semantic vectors corresponding to the candidate arguments. ; The semantic vector of the candidate arguments after being encoded by the BERT pre-trained model; For the first One recommended candidate argument. For argument confidence weights in recommendation algorithms, As candidate arguments, As another candidate argument; Ultimately, the argument role is obtained. of Recommended results ; Recommendation results The expression is as follows: .

[0064] This sorting method ensures that the recommended list includes both high-confidence candidates and covers different forms of expression, thus improving the flexibility of manual selection.

[0065] For example: Time: (June 17, 2024, evening, 0.9); (June 17, evening, 0.4); Winning teams: (Team A, 0.95); (Team B, 0.11); Losers: (Team B, 0.92); (Team A, 0.06); Score: (106 to 88, 0.95); (4 to 1, 0.26).

[0066] Example 2: As a further embodiment of the present invention, the present invention also provides an electronic device, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0067] In practical use, users can interact with servers, which are also electronic devices, via a network to receive or send messages. Terminal devices are generally various electronic devices equipped with a display and used through a human-computer interface, including but not limited to smartphones, tablets, laptops, and desktop computers. Various specific application software can be installed on these terminal devices as needed, including but not limited to web browsers, instant messaging software, social media platforms, and shopping apps.

[0068] Furthermore, a server is a network service provider that offers various services, such as processing input text received from terminal devices. Event trigger words Event Type and the set of roles defined for this type of event. A backend server provides corresponding computing services to process the event argument recommendation method, calculate a high-quality, high-coverage unified candidate list, and finally return it to the terminal device.

[0069] Example 3: As a further embodiment of the present invention, the present invention also provides a storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the event argument recommendation method as described above.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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. An event argument recommendation method based on multi-model fusion, characterized in that, Includes the following steps: Step 1: Let the input text be... Event trigger words Event Type The set of roles defined for any type of event is and each argument role ; Step 2: Assemble the characters Each argument role in The input is fed into multiple argument extraction sub-models with different technical approaches, and multiple predicted arguments and their corresponding confidence scores are obtained respectively. Step 3: Combine multiple predicted arguments and their corresponding confidence scores to form a candidate argument set. ; Step 4: Set of candidate arguments Any argument role in The output includes The set of recommendation arguments for each result; Multiple argument extraction sub-models with different technical approaches include question-answering guided extraction model, template filling extraction model and dialogue generation model; The specific process of generating predicted arguments using a question-and-answer guided extraction model is as follows: S2.11 for each argument role Constructing its natural language problem ; S2.12, Natural Language Problems With input text The data is concatenated together and used as input to the pre-trained model in the question-answering guided extraction model. ; S2.13, based on Obtain the context representation ; S2.14, Context-based representation Construct the head pointer Sum of tail pointers And based on the head pointer Sum of tail pointers Determine any argument role starting position and end position ; S2.15, Set the candidate text fragments of the event arguments as follows: Based on any argument role starting position and end position Calculate the set of candidate text fragments Combined score ; S2.16, Joint score based on candidate text fragments Select the highest-scoring legitimate text segment as the model's prediction argument. ; S2.17 Predicted arguments for all candidate text segments Perform Softmax normalization to obtain its confidence score. ; The specific process of generating predictive arguments using a template-filling extraction model is as follows: S2.21 Input Construction; The construction includes fill-in-the-blank natural language prompt templates for all roles. ; S2.

22. Using a model with an Encoder-Decoder architecture, the input text is processed... This serves as the input and prompt template for the Encoder part of the model. As input to the Decoder part of this model, an event-oriented context representation is obtained. and context-oriented hints ; S2.23, Define the role of each argument. have One slot; S2.24 Context-Oriented Hint Representation Get each argument role The Role characteristics of each slot ;in, ; S2.25, From the perspective of character traits The role of each argument is derived from the derivation. The Text snippet selector for each slot ; Use text fragment selector From the context of event-oriented representation Extract a corresponding event argument; S2.

26. Repeat steps S2.24 and S2.25 to obtain the role of each argument. The event arguments corresponding to all slots are used to define the resulting event arguments as a set of candidate text fragments. ; For the candidate text fragment set Score for each text segment Perform calculations; S2.27, Scoring based on each text segment The text segment with the highest score was selected as the prediction argument. ; S2.28 Predicted arguments for all candidate text segments Perform Softmax normalization to obtain the predicted arguments. confidence level .

2. The event argument recommendation method based on multi-model fusion according to claim 1, characterized in that, The specific process of generating predicted arguments using a dialogue-based generative model is as follows: S2.31, Constructor Instruction Hints ,in, These are the prompt words input to the generative model; S2.32, Define the role of each argument. The generated prediction arguments are ; Based on the intermediate information generated during decoding, the generation probability is... ; S2.33, Based on the generation probability Calculate the role of each argument Includes the average log probability of each word and assigns each argument role The average log probability of the included word as the generation confidence. .

3. The event argument recommendation method based on multi-model fusion according to claim 2, characterized in that, The specific method for step three is as follows: For each argument role Its candidate argument set ,in, For the corresponding text fragment, For model confidence, As the source of the model, These are the aligned candidate arguments.

4. The event argument recommendation method based on multi-model fusion according to claim 3, characterized in that, The specific process of step four is as follows: S4.

1. Let any one of the argument extraction sub-models from multiple different technical approaches be... The extreme value standardization method is used for any sub-model. Normalization is performed to obtain the normalized confidence level. ; S4.2, Based on normalized confidence level Calculate each aligned candidate argument Fusion score ; S4.3 From the set of candidate arguments The Top-K recommended items are selected, and the maximum marginal relevance algorithm is used to balance relevance and diversity to obtain the argument roles. of Recommended results .

5. The event argument recommendation method based on multi-model fusion according to claim 4, characterized in that, Obtaining the role of the argument of Recommended results The specific process is as follows: Recommendation results Perform initialization: ; in, It is an empty set. This is the set of candidate arguments that were not recommended. Iteration to ; ； ； in, This is a function for calculating the similarity between two candidate arguments, obtained by calculating the cosine similarity of the semantic vectors corresponding to the candidate arguments; , The semantic vector of the candidate arguments after being encoded by the BERT pre-trained model; For the first One recommended candidate argument. For argument confidence weights in recommendation algorithms, As candidate arguments, As another candidate argument; Based on the argument role of Recommended results ; 。 6. An electronic device, characterized in that, It includes a memory, one or more processors, and one or more programs stored in the memory, said one or more programs including instructions for executing the event argument recommendation method as described in any one of claims 1-5.

7. A storage medium, characterized in that, It includes one or more programs that are executed by one or more processors of an electronic device, the one or more programs including instructions for performing the event argument recommendation method as described in any one of claims 1-5.

Images