Evidence-guided generation and original text positioning correction method and system for zero-shot event argument extraction

Abstract

The application relates to the technical field of event argument extraction, and discloses an evidence guidance generation and original text positioning correction method and system for zero-sample event argument extraction, wherein the method comprises the following steps: obtaining event text data to be processed and formatting the event text data into event extraction index instances; constructing a candidate argument text segment set based on the instances; performing function type labeling, role constraint mapping and evidence item encapsulation on the candidate segments to obtain a structured evidence set; constructing evidence guidance prompts based on the structured evidence set and generating an intermediate structured generation result in a JSON format; then performing original text segment alignment, role legality verification, repeated record deletion and segment boundary correction; and outputting event argument extraction results. The application solves the problems of difficulty in returning to the original text, role boundary crossing and unstable segment boundaries in the generation of arguments in a zero-sample scenario, and improves the localizability and reviewability of argument extraction results.

Description

Technical Field

[0001] This invention relates to the field of event argument extraction technology, and in particular to a method and system for evidence-guided generation and original text location verification and correction for zero-sample event argument extraction. Background Technology

[0002] Currently, event argument extraction is a key task in constructing structured event representations from unstructured text. It typically involves identifying consecutive text fragments related to the current event and assigning them to corresponding event roles, given constraints such as event trigger words, event types, and event roles. This task is widely used in scenarios such as event understanding, knowledge base construction, intelligence analysis, and information retrieval.

[0003] Existing event argument extraction methods mostly rely on supervised learning, requiring the construction of a large number of labeled samples for specific event types and role constraints. Due to the fine-grained semantics of event roles, the high requirements for argument boundary annotation, and the continuous changes in event types and role constraints in professional domains, supervised methods struggle to maintain stability in cross-domain and unseen role constraint scenarios.

[0004] Zero-shot event argument extraction attempts to extract arguments without using target task-labeled samples or updating parameters. In this scenario, the model needs to satisfy three constraints simultaneously: the extracted result must be locatable as a continuous text segment in the original text; the role label must belong to the legitimate role set of the given event type; and the argument semantics must be consistent with the given event trigger word. Existing prompt-based or generation-based solutions, while possessing good generalization ability, are prone to problems such as phantom arguments, imprecise boundaries, invalid role labels, and duplicate extraction.

[0005] Furthermore, event arguments are not limited to named entities; they may also include temporal expressions, numerical expressions, tool phrases, descriptive clauses, and resultative phrases. Relying solely on a fixed set of candidate entities can easily lead to the omission of non-entity arguments; relying entirely on freely generated arguments makes it difficult to guarantee consistency between original text location and role constraints. Therefore, it is necessary to provide a zero-shot event argument extraction method that can balance open coverage and structured validation capabilities. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the present invention aims to provide an evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction, thereby solving the problem that existing zero-sample event argument extraction methods struggle to balance generation flexibility, original text location accuracy, and consistency of role constraints.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides an evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction.

[0008] The evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction includes: Step S10: Obtain the event text data to be processed, and perform event instance formatting on the event text data to be processed using the lexical numbering and role list extraction mechanism, and output the event extraction index instance; Step S20: Based on the event extraction index instance, perform candidate argument text fragment construction using original text fragment scanning and open evidence discovery prompting mechanism, and output a set of candidate argument text fragments; Step S30: Based on the candidate argument text fragment set, perform structured evidence construction using functional type labeling, role constraint mapping, and evidence item encapsulation mechanisms, and output a structured evidence set; Step S40: Based on the structured evidence set, the argument role pair generation mechanism is used to construct evidence-guided prompts and generate JSON format constraints, and the intermediate structured generation result is output. Step S50: Based on the intermediate structured generation result, the positioning verification and correction are performed using the original text fragment alignment, role legality verification, duplicate record deletion and fragment boundary correction mechanism, and the event argument extraction result is output.

[0009] Preferably, in step S10, the steps of acquiring the event text data to be processed, performing event instance formatting on the event text data to be processed using a lexical numbering and role list extraction mechanism, and outputting the event extraction index instance specifically include: Step S101: Obtain the input context from the text data of the event to be processed, number the input context according to the word order, and output the indexed word sequence; Step S102: Read the event trigger word text fragment, event trigger word index and event type from the event text data to be processed, and encapsulate the event trigger word text fragment, the event trigger word index and the event type into an event mention field; Step S103: Read the preset event role constraints from the event text data to be processed, and extract the list of allowed roles corresponding to the event type; Step S104: Encapsulate the indexed term sequence, the event mention field, and the allowed role list into the event extraction index instance.

[0010] Preferably, step S20, which involves constructing candidate argument text fragments based on the event extraction index instance using original text fragment scanning and an open evidence discovery hint mechanism, and outputting a set of candidate argument text fragments, specifically includes: Step S201: Read the indexed word sequence and the event mention field from the event extraction index instance, perform named entity recognition, time expression recognition, numerical expression recognition and noun phrase extraction on the indexed word sequence, and output the first candidate fragment list; Step S202: Write the event extraction index instance into a preset open evidence discovery prompt word, and the large language model returns a second candidate fragment list containing only the original text fragment, evidence category, clue description and confidence level; Step S203: Perform original text occurrence verification on the first candidate segment list and the second candidate segment list, and delete candidate segment records that do not appear in the input context; Step S204: Extract the candidate fragment records that pass the original text occurrence verification by the minimum independently comprehensible boundary, and record the start and end lexical offsets of the extracted candidate fragment records in the indexed lexical sequence, and output the candidate argument text fragment set.

[0011] Preferably, step S30, which involves constructing structured evidence based on the candidate argument text fragment set using functional type labeling, role constraint mapping, and evidence item encapsulation mechanisms, and outputting a structured evidence set, specifically includes: Step S301: Assign at least one broad argument function type to each candidate argument text fragment in the candidate argument text fragment set, the broad argument function type including at least one of participant, location, time, tool, affected entity, content, value, result and others; Step S302: Read the allowed role list from the event extraction index instance, perform compatible role mapping on the candidate argument text fragments with assigned broad argument function types according to the allowed role list, and output the compatible role set corresponding to each candidate argument text fragment; Step S303: Record auxiliary metadata for each candidate argument text fragment, wherein the auxiliary metadata includes at least one of evidence source, confidence level, lexical distance to the event trigger word, and matching support information; Step S304: Encapsulate each candidate argument text fragment, the lexical offset of the candidate argument text fragment, the broad argument function type, the compatible role set, and the auxiliary metadata into an evidence item, and denote the input context as... Let the list of allowed roles be denoted as The evidence item is represented as follows: in, Indicates the first One piece of evidence, Indicates the first A candidate argument text fragment, Indicates the first The initial lexical offset of each candidate argument text fragment. Indicates the first End-word offset of candidate argument text fragments Indicates the first A set of broad argument function types corresponding to each candidate argument text fragment. Indicates the first A set of compatible roles corresponding to each candidate argument text fragment, and satisfying the following conditions: , Indicates the first Auxiliary metadata corresponding to each candidate argument text fragment; Step S305: The structured evidence set is composed of multiple evidence items.

[0012] Preferably, step S40, which involves generating argument / role pairs based on the structured evidence set using an evidence-guided prompting construction and JSON format constraint generation mechanism, and outputting intermediate structured generation results, specifically includes: Step S401: Merge identical text fragments and normalize identical text fragments in the evidence items of the structured evidence set, and output a deduplicated evidence set; Step S402: Extract the evidence item corresponding to each allowed role from the deduplicated evidence set according to the preset number of role evidence to retain, and construct a compact evidence table; Step S403: Extract the top-ranked evidence items from the deduplicated evidence set according to the preset global evidence retention quantity, and construct a global evidence list; Step S404: Write the event extraction index instance, the compact evidence table, the global evidence list, and the preset output format constraints into the evidence guiding argument generation prompt words; Step S405: Input the evidence-guided argument generation prompt into the large language model that follows the instructions, so that the large language model generates candidate argument role records according to the preset output format constraints, and outputs the intermediate structured generation result; wherein, the candidate argument role record includes an argument text field, a role field, and an evidence usage field, and the compact evidence table and the global evidence list are only used as original text location guidance information, and are not used as a fixed candidate set to limit the generation range of the large language model.

[0013] Preferably, step S50, which involves performing location verification and correction based on the intermediate structured generation result using mechanisms such as original text fragment alignment, role validity verification, duplicate record deletion, and fragment boundary correction, and outputting the event argument extraction result, specifically includes: Step S501: Read each candidate argument role record from the intermediate structured generation result, perform exact string matching on the argument text field in the candidate argument role record in the input context, and output the first original text aligned record; Step S502: When the first original text alignment record is not obtained, perform case folding, punctuation normalization and whitespace character normalization on the argument text field, and perform string matching between the normalized argument text field and the input context to output the second original text alignment record; Step S503: When the first original text alignment record or the second original text alignment record includes multiple candidate alignment positions, the candidate alignment position with evidence item support is selected first; when multiple candidate alignment positions are supported by evidence items, the candidate alignment position with the smallest lexical distance to the event trigger word index is selected. Step S504: Compare the role field in the candidate argument role record with the allowed role list, delete candidate argument role records whose role field is not included in the allowed role list, and output a set of valid role records; Step S505: Deduplicate candidate argument role records in the legal role record set that have the same role and the same or overlapping text fragments, and retain candidate argument role records that meet the preset retention conditions in terms of original text alignment and evidence support. Step S506: Perform fragment boundary correction on the candidate argument role records after deduplication, and output the event argument extraction result. The event argument extraction result includes argument text fragments that can be located in the input context and their corresponding legal roles.

[0014] Preferably, step S506, which involves performing fragment boundary correction on the candidate argument role records after deduplication, specifically includes: Read the aligned text fragments and role fields from the candidate argument role records after deduplication; Extract evidence items that are the same as or compatible with the role field from the structured evidence set, and output the evidence items to be compared; When the aligned text fragment satisfies the text fragment inclusion relationship with the candidate argument text fragment in the evidence item to be compared, the start word offset and end word offset of the aligned text fragment are replaced with the start word offset and end word offset of the candidate argument text fragment in the evidence item to be compared. When the aligned text fragment and the candidate argument text fragment in the evidence item to be compared do not satisfy the text fragment inclusion relationship, but the word overlap ratio of the two reaches the preset boundary correction threshold, the start word offset and end word offset of the aligned text fragment are replaced by the start word offset and end word offset of the candidate argument text fragment in the evidence item to be compared. When the preset recovery switch is enabled, text fragments that have been located in the input context, have a role in the allowed role list, and have evidence item support will be restored as candidate argument role records; The restored candidate argument role records are subjected to role validity checks and duplicate record deletion again, and the restored candidate argument role records are not generated by additional large language model generation calls.

[0015] This invention also provides an evidence-guided generation and original text location verification and correction system for zero-sample event argument extraction, including: The event instance formatting unit is used to acquire event text data to be processed, perform event instance formatting on the event text data to be processed using a lexical numbering and role list extraction mechanism, and output an event extraction index instance. The candidate fragment construction unit is used to construct candidate argument text fragments based on the event extraction index instance using original text fragment scanning and open evidence discovery prompting mechanism, and outputs a set of candidate argument text fragments. The structured evidence construction unit is used to perform structured evidence construction based on the candidate argument text fragment set using functional type labeling, role constraint mapping and evidence item encapsulation mechanisms, and output a structured evidence set; The evidence-guided generation unit is used to generate argument-role pairs based on the structured evidence set using evidence-guided prompts and a JSON format constraint generation mechanism, and output intermediate structured generation results. The original text positioning verification and correction unit is used to perform positioning verification and correction based on the intermediate structured generation result by means of original text fragment alignment, role legality verification, duplicate record deletion and fragment boundary correction mechanism, and output the event argument extraction result.

[0016] The present invention also provides an evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction. The evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction includes: a memory, a processor, and an evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction stored in the memory and executable on the processor. When the evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction is executed by the processor, it implements the above-described method.

[0017] The present invention also provides a computer program product, the computer program product including an evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction, the evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction implementing the above method when executed by a processor.

[0018] The beneficial effects of this invention are as follows: 1. This invention provides fragment-level, role-constraint-aware soft guidance to the pre-defined generation model by constructing a structured evidence set before generation, thereby improving the coverage of heterogeneous arguments such as named entities, time, numerical values, phrases, and clauses in zero-shot scenarios. By performing deterministic original text location verification and boundary correction after generation, it can reduce illusory arguments, illegal roles, and boundary drift. By using evidence as advisory guidance rather than hard candidate constraints, it ensures that the final output can fall back to the original continuous text fragments and the set of legitimate roles while retaining open generation capabilities.

[0019] 2. This invention encapsulates candidate argument text fragments into structured evidence items and configures them with broad argument function types, compatible role sets, lexical offsets, and auxiliary metadata. This transforms the evidence from a loose list of texts into structured objects with role constraints and source descriptions. This structured evidence set provides traceable supporting material for evidence-guided generation, reducing false positives caused by confusion in candidate role mapping.

[0020] 3. This invention constructs a compact evidence table and a global evidence list, and combines JSON format constraints to generate intermediate structured results, thus subjecting the large language model to dual constraints of evidence items and output format during open generation. This approach differs from directly limiting a fixed candidate set, preserving the generative model's flexibility in handling descriptive arguments while reducing inconsistencies between role fields, argument text fields, and evidence usage fields. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the first embodiment of the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction according to the present invention.

[0022] Figure 2 This is a schematic diagram of the event argument extraction task in the first embodiment of the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction of the present invention.

[0023] Figure 3 This is a schematic diagram of the evidence-guided generation and original text location verification framework for the first embodiment of the evidence-guided generation and original text location verification method for zero-sample event argument extraction of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction according to the present invention. The first embodiment of the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction according to the present invention is presented.

[0027] In the first embodiment, the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction includes: Step S10: Obtain the event text data to be processed, and perform event instance formatting on the event text data to be processed using the lexical numbering and role list extraction mechanism, and output the event extraction index instance; The event text data to be processed refers to the original text and its event constraint information that enters the argument extraction process. This typically includes the input context, event trigger word text fragments, event trigger word indexes, event types, and pre-defined event role constraints corresponding to those event types. The lexical numbering and role list extraction mechanism does not simply feed the text into the model; instead, it first establishes stable numbers for the input context according to the original lexical order, then encapsulates the trigger word text, trigger word index, and event type into an event mention field, while simultaneously extracting the list of allowed roles from the role constraints. An event extraction index instance consists of an indexed lexical sequence, an event mention field, and a list of allowed roles. Each subsequent candidate fragment, evidence item, and final argument can be traced back to the lexical position and role range within this instance.

[0028] This formatting process first establishes a unified input coordinate system for zero-sample event argument extraction. Since the argument text generated by the large language model may be rewritten, if the front-end does not retain the original word IDs, it is difficult to determine whether the generated string truly corresponds to the original text segment. Step S10 binds the original text order, trigger word position, and allowed roles simultaneously; step S20 records the start and end word offsets when candidate segments are found; and step S50 uses the same set of IDs for boundary correction when aligning the original text segments. Each step no longer uses inconsistent text segmentation standards.

[0029] Traditional zero-shot extraction often directly concatenates the input context, trigger words, and role descriptions into prompt words. While simple to implement, it lacks a reusable structured index. When the text contains multiple dates, locations, or organization names, it's difficult to determine the corresponding original text location after the model outputs an argument. Furthermore, when the role set changes, the role constraints in the old prompt words are difficult to reuse in subsequent validation steps. Step S10 preserves the trigger word index and allowed role list through event extraction index instances, ensuring that subsequent candidate discovery, evidence encapsulation, and role legitimacy verification all revolve around the same event instance, reducing cross-step information loss.

[0030] For example, in the news text "At 2:30 PM on March 15, 2025, a passenger plane carrying 132 passengers crashed in the eastern mountains, causing significant casualties," the trigger word is "crash," and the event type is air crash. Step S10 will assign word element numbers to segments such as "March 15, 2025," "2:30 PM," "132 passengers," and "eastern mountains" according to word element order, and write the word element position of "crash" into the event mention field. Simultaneously, it will read the allowed time, location, passenger capacity, and affected entities for the air crash event. Thus, when "132 passengers" are subsequently discovered, not only will their text content be known, but also their word element range in the original text and the possible role constraints they may be subject to.

[0031] Step S20: Based on the event extraction index instance, perform candidate argument text fragment construction using original text fragment scanning and open evidence discovery prompting mechanism, and output a set of candidate argument text fragments; The original text fragment scanning and open evidence discovery prompt mechanism refers to the process of generating candidate fragments from three levels: rule recognition, shallow language analysis, and open model discovery, based on the indexed lexical sequences and event mention fields in the event extraction index instances. Specifically, the system performs named entity recognition, temporal expression recognition, numerical expression recognition, and noun phrase extraction on the indexed lexical sequences to obtain a first candidate fragment list. Then, the event extraction index instances are written into preset open evidence discovery prompt words, and the large language model returns a second candidate fragment list containing only the original text fragment, evidence category, clue description, and confidence level. After merging the two lists, each candidate fragment is checked to ensure it actually appears in the input context, and the fragments that pass the verification are truncated to the smallest independently comprehensible boundary, while the start and end lexical offsets are recorded.

[0032] The technical effect of this step is to limit candidate argument discovery to verifiable fragments in the original text. Named entities, temporal expressions, and numerical expressions can cover well-structured arguments, noun phrase extraction can supplement descriptive objects, and open-ended evidence discovery hints are used to capture phrases or clauses that do not conform to fixed entity categories but are related to event roles. Original text occurrence verification blocks the path for the model to rewrite or generate non-existent fragments on its own, and minimum independently comprehensible boundary truncation avoids candidate fragments that are too long, too short, or contain only modifiers. The output set of candidate argument text fragments retains both text content and lexical offsets, providing clear objects for the function type labeling and role constraint mapping in step S30.

[0033] Traditional candidate discovery methods, relying solely on named entity recognition, are prone to missing numerical and outcome phrases such as "132 passengers" and "causing significant casualties." Conversely, relying entirely on open generation from large language models can easily result in generalized statements that are not continuous fragments of the original text, such as "there were many people on board" and "the accident occurred in the afternoon." Step S20 combines rule scanning and open discovery, but uses original text occurrence verification as a hard constraint, ensuring the model can only contribute evidence fragments that can be traced back to the input context. This reduces the risk of missing non-entity arguments and introducing illusory fragments, even without labeled samples.

[0034] Using the air crash news example, rule scanning can identify "2:30 PM on March 15, 2025," "132 people," and "eastern mountainous area." Noun phrase extraction yields "a passenger plane carrying 132 passengers," and open-ended evidence discovery suggests "significant casualties" as a possible outcome clue. The system checks whether these fragments appear word-for-word in the input context. If the model returns "near the eastern mountainous area" but the original text only mentions "eastern mountainous area," then "near the eastern mountainous area" will not be directly considered a candidate fragment. If the boundary of "a passenger plane carrying 132 passengers" is too long, the system can truncate it into a candidate fragment that can still be understood independently and record its lexical offset. Each record in the final set can be accurately referenced in subsequent steps.

[0035] Step S30: Based on the candidate argument text fragment set, perform structured evidence construction using functional type labeling, role constraint mapping, and evidence item encapsulation mechanisms, and output a structured evidence set; The function type labeling, role constraint mapping, and evidence item encapsulation mechanism refers to converting the candidate argument text fragments obtained in step S20 into structured evidence items that can be used by the generative model. Each candidate fragment is first assigned a broad argument function type, which can cover categories such as participants, locations, times, tools, affected entities, content, values, results, and others. Then, the system reads the list of allowed roles from the event extraction index instance and maps the candidate fragments with assigned function types to one or more compatible roles. Next, auxiliary metadata such as evidence source, confidence level, lexical distance to the event trigger word, and matching support information are recorded for the candidate fragment. After completion, each evidence item contains the candidate fragment text, lexical offset, broad argument function type, set of compatible roles, and auxiliary metadata. Multiple evidence items form a structured evidence set.

[0036] This step upgrades candidate fragments from plain text into evidence objects with role constraints and source descriptions. Broad argument function types provide coarse-grained semantic classification, identifying "2:30 PM" as a time-related fragment and "132 passengers" as a numerical or affected entity-related fragment. A compatible role set, combined with a list of allowed roles, restricts candidate fragments to the range of roles that the current event type can accept. Auxiliary metadata records why the evidence is retained and its relative position to the trigger word. Thus, when constructing evidence guidance prompts in step S40, it is not necessary to deal with scattered candidate fragments; instead, a structured evidence set with role candidates and offset information can be directly used.

[0037] Traditional methods often directly provide the candidate fragment list to the generation model, allowing the model to determine which fragment corresponds to which role. This approach is prone to role confusion when there are many candidate fragments; for example, an organization name might be mistaken for a participant, or a news source might be mistaken for the event subject. Step S30 addresses this by first labeling broad functional types and then performing a compatibility mapping with the allowed role list, thus restricting or excluding fragments that cannot possibly assume the current event role. Because the structured evidence retains lexical offsets and auxiliary metadata, subsequent generation results referencing this evidence can trace its original text location and supporting sources, reducing the uninterpretability of black-box generation.

[0038] For example, although the candidate fragment "Xinhua News Agency" is a named entity, its functional type is closer to the information source and may not necessarily belong to the core arguments of the air crash event; after the candidate fragment "2:30 pm on March 15, 2025" is labeled as a time-related fragment, it can be mapped to the role of the time of occurrence; after "eastern mountainous area" is labeled as a location-related fragment, it can be mapped to the role of the location of occurrence; "132 passengers" has both numerical and affected entity attributes, and can be mapped to passenger volume or affected entity roles. The system encapsulates the text, offset, type, compatible roles, and source confidence of each fragment into evidence items, enabling step S40 to organize evidence around roles, rather than letting the generative model search in the unstructured candidate pool on its own.

[0039] Step S40: Based on the structured evidence set, the argument role pair generation mechanism is used to construct evidence-guided prompts and generate JSON format constraints, and the intermediate structured generation result is output. The evidence-guided prompt construction and JSON format constraint generation mechanism involves first merging identical text fragments and normalizing the merging of identical text fragments in the structured evidence set, outputting a deduplicated evidence set; then, according to the preset role evidence retention quantity, extracting corresponding evidence items for each allowed role to construct a compact evidence table; simultaneously, extracting globally prioritized evidence items according to the preset global evidence retention quantity to construct a global evidence list. Subsequently, the system writes the event extraction index instance, compact evidence table, global evidence list, and preset output format constraints into the evidence-guided argument generation prompt, and inputs it into a large language model that follows the instructions, causing it to generate candidate argument role records containing argument text fields, role fields, and evidence usage fields, forming an intermediate structured generation result.

[0040] The technical advantage of this step lies in maintaining a balance between open generation capabilities and evidence constraints in the large language model. The compact evidence table ensures that each allowed role has priority evidence items to refer to, while the global evidence list retains important fragments shared across roles, preventing certain evidence from being completely excluded due to initial role-based bias. JSON format constraints allow the model output to have fixed fields, facilitating direct reading of argument text, roles, and evidence usage in step S50. Since the compact evidence table and global evidence list are merely text-based location guidance information rather than a fixed candidate set, the model can still generate descriptive arguments based on context, but must explain the evidence used in a structured manner.

[0041] Traditional prompt-based extraction typically inputs the full text and character descriptions into the model all at once, requiring the model to directly output arguments. This approach is prone to ignoring precise numerical values, confusing similar times and locations, or exhibiting unstable output formats. Another type of extraction method only allows selection from a candidate set; while providing stable localization, it lacks flexibility when dealing with descriptive clauses or complex noun phrases. Step S40 uses structured evidence as soft guidance and constrains the output format with JSON fields, enabling the model to both utilize its generative capabilities to handle open arguments and establish connections between each generated record and the evidence item. This intermediate structured generation result provides a clear entry point for subsequent validation.

[0042] For example, regarding the role of the time of the air crash, the compact evidence table could list "2:30 PM on March 15, 2025"; regarding the role of the location, it could list "eastern mountainous area"; the global evidence list could also include "aircraft carrying 132 passengers" and "major casualties". When the large language model outputs under cue constraints, it needs to provide the argument text, role field, and evidence items used in JSON record form. If the model attempts to output "around 2:30 PM", although the text is semantically close, the evidence usage field will point to "2:30 PM" in the original text, allowing step S50 to detect the expression difference between the generated text and the original text fragment and perform localization correction.

[0043] Step S50: Based on the intermediate structured generation result, the positioning verification and correction are performed using the original text fragment alignment, role legality verification, duplicate record deletion and fragment boundary correction mechanism, and the event argument extraction result is output.

[0044] The mechanism for original text fragment alignment, role validity verification, duplicate record deletion, and fragment boundary correction involves reading candidate argument role records one by one from the intermediate structured generation results. First, precise string matching is performed on the argument text field to obtain the first original text aligned record. If no match is found, the argument text undergoes case folding, punctuation normalization, and whitespace normalization before being matched against the input context to obtain the second original text aligned record. If multiple candidate alignment positions exist, the system prioritizes the position supported by evidence. If multiple positions are supported by evidence, the position with the smallest distance to the event trigger word index term is selected. Subsequently, the role field must be compared with the list of allowed roles; records not included in the list are deleted. Records with the same role and identical or overlapping text are deduplicated, retaining only those that meet the conditions for original text alignment and evidence support. Finally, fragment boundary correction is performed on the retained records, outputting the argument text fragments that can be located in the input context and their valid roles.

[0045] This step transforms the open-source generated results into actionable event argument extraction results. Exact matching and normalized matching jointly handle differences in capitalization, punctuation, and whitespace in the model output; multi-position selection rules utilize evidence item support and trigger word distance to resolve ambiguity when the same fragment appears multiple times; role validity checks prevent the model from generating roles outside the allowed role list; duplicate record deletion reduces the multiple outputs of the same argument; and boundary correction adjusts potentially generated general phrases into continuous fragments from the original text. After these processes, the final result not only contains arguments and roles but also clearly points to specific locations within the input context.

[0046] Traditional large language models often produce outputs that appear semantically correct, but they cannot guarantee that each argument can be located word-for-word in the original text, nor can they guarantee that role names are consistent with predefined role constraints. In news scenarios involving numbers, times, and locations, even a single digit or a boundary word can alter downstream analysis results. Step S50, by pulling the generated records back into the original text and filtering them based on supporting evidence and role legitimacy, ensures that the model's free generation is constrained by a verifiable chain of evidence. Compared to simply requiring "fidelity to the original text" in prompts, this step provides a practically executable verification path, capable of identifying and correcting model rewriting, duplication, or out-of-bounds outputs.

[0047] For example, the model outputs the argument "133 passengers" in the intermediate structured generation result, with the role being passenger capacity, while the actual input context states "132 passengers". Step S50 first performs an exact match, but cannot find "133 passengers" in the original text; after normalization, it still cannot match; the system then checks the evidence item claimed by the record and finds that the evidence item corresponds to the original text segment "132 passengers", the two numbers are inconsistent, so the record cannot be directly retained. If the model outputs "passengers" and points to the evidence item "132 passengers", the system can correct the boundary to the complete segment "132 passengers" based on the start and end lexical offsets of the evidence item. The final output event argument extraction result retains the correct role and remains consistent with the continuous segment of the original text.

[0048] For example, such as Figure 2As shown in the diagram, this is a schematic diagram of an event argument extraction task. It illustrates the process of identifying arguments related to the current event mention in instance text and mapping them to the corresponding event roles, given event trigger words, event types, and role constraints. The instance text in the diagram describes a "thug" dragging a family of three out and killing them with a "machete" and a "spear" before fleeing the "area." Here, "kill" is the event trigger word, and the event type is "conflict.attack." Based on the allowed roles corresponding to this event type, the original text fragment "thug" is mapped to the attacker, "they" is mapped to the target, "machete" and "spear" are mapped to weapons, and "area" is mapped to a location. Therefore, the arguments to be extracted should be text fragments that can be continuously located in the original text. The same event role can correspond to multiple argument fragments, and the extraction results should also satisfy the role range limited by the event type.

[0049] like Figure 3 As shown in the figure, this invention illustrates the overall processing framework consisting of input, evidence memory construction, generation, and verification. The input phase acquires instance text, event trigger words, and event types, and reads allowed roles, entity-role compatibility, and other pattern information from event patterns, providing constraints for subsequent evidence projection and role legitimacy verification. The evidence memory construction phase first performs open evidence discovery, extracting potential argument fragments such as "thugs," "they," "machete," "spear," and "region" from the original text, and labeling them as broad functional types such as participants, weapons, or locations. Then, based on event patterns, aligned projection is performed, projecting participants to attackers or targets, weapons to weapons, and locations to places. Finally, duplicate or overlapping evidence is merged to form a structured evidence memory containing original text fragments, functional types, and compatible role information. The generation and verification phase inputs instance text, trigger words, event type, event pattern, and evidence memory into the argument generation model to obtain the original argument results. Then, it sequentially performs original text verification, fragment repair, duplicate result deletion, and evidence support restoration to delete results that cannot be located in the original text or have illegitimate roles, correct argument boundaries, and restore legal arguments with evidence support. Finally, it outputs argument role pairs that can be located in the original text and conform to role constraints, such as attacker and "thugs", target and "them", weapon and "machete" or "spear", location and "region".

[0050] Example 2: Furthermore, the evidence-guided generation and original text location verification and correction system for zero-sample event argument extraction provided by this invention employs the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction described in the above embodiments, and can solve the technical problems of evidence-guided generation and original text location verification and correction for zero-sample event argument extraction. The beneficial effects of the evidence-guided generation and original text location verification and correction system for zero-sample event argument extraction provided by this invention are the same as those of the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction provided in the above embodiments, and other technical features of the evidence-guided generation and original text location verification and correction system for zero-sample event argument extraction are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0051] Example 3: This invention provides an evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which are then executed to enable the processor to perform the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction described in Example 1. The evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction in this invention can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction is merely an example and should not limit the functionality or scope of the invention. An evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in read-only memory or a program loaded from a storage device into random access memory. The random access memory also stores various programs and data required for the operation of the evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction. The processing unit, read-only memory, and random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction to communicate wirelessly or wiredly with other devices to exchange data. While evidence-guided generation and original text location verification and correction devices for zero-sample event argument extraction with various systems have been described, it should be understood that implementation of or possession of all the described systems is not required. Alternatively, more or fewer systems may be implemented.

[0052] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described above. The computer program product provided by this invention can solve the technical problems of evidence-guided generation and original text location verification and correction for zero-sample event argument extraction. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction provided in the above embodiments, and will not be repeated here.

[0053] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0054] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0055] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for evidence-guided generation and original text location verification and correction for zero-sample event argument extraction, characterized in that, The methods include: Step S10: Obtain the event text data to be processed, and perform event instance formatting on the event text data to be processed using the lexical numbering and role list extraction mechanism, and output the event extraction index instance; Step S20: Based on the event extraction index instance, perform candidate argument text fragment construction using original text fragment scanning and open evidence discovery prompting mechanism, and output a set of candidate argument text fragments; Step S30: Based on the candidate argument text fragment set, perform structured evidence construction using functional type labeling, role constraint mapping, and evidence item encapsulation mechanisms, and output a structured evidence set; Step S40: Based on the structured evidence set, the argument role pair generation mechanism is used to construct evidence-guided prompts and generate JSON format constraints, and the intermediate structured generation result is output. Step S50: Based on the intermediate structured generation result, the positioning verification and correction are performed using the original text fragment alignment, role legality verification, duplicate record deletion and fragment boundary correction mechanism, and the event argument extraction result is output.

2. The evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in claim 1, characterized in that, In step S10, the process of acquiring the event text data to be processed, performing event instance formatting on the event text data using a lexical numbering and role list extraction mechanism, and outputting the event extraction index instance specifically includes: Step S101: Obtain the input context from the text data of the event to be processed, number the input context according to the word order, and output the indexed word sequence; Step S102: Read the event trigger word text fragment, event trigger word index and event type from the event text data to be processed, and encapsulate the event trigger word text fragment, the event trigger word index and the event type into an event mention field; Step S103: Read the preset event role constraints from the event text data to be processed, and extract the list of allowed roles corresponding to the event type; Step S104: Encapsulate the indexed term sequence, the event mention field, and the allowed role list into the event extraction index instance.

3. The evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in claim 2, characterized in that, Step S20, which involves constructing candidate argument text fragments based on the event extraction index instance using original text fragment scanning and an open evidence discovery hint mechanism, and outputting a set of candidate argument text fragments, specifically includes: Step S201: Read the indexed word sequence and the event mention field from the event extraction index instance, perform named entity recognition, time expression recognition, numerical expression recognition and noun phrase extraction on the indexed word sequence, and output the first candidate fragment list; Step S202: Write the event extraction index instance into a preset open evidence discovery prompt word, and the large language model returns a second candidate fragment list containing only the original text fragment, evidence category, clue description and confidence level; Step S203: Perform original text occurrence verification on the first candidate segment list and the second candidate segment list, and delete candidate segment records that do not appear in the input context; Step S204: Extract the candidate fragment records that pass the original text occurrence verification by the minimum independently comprehensible boundary, and record the start and end lexical offsets of the extracted candidate fragment records in the indexed lexical sequence, and output the candidate argument text fragment set.

4. The evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in claim 3, characterized in that, Step S30, which involves constructing structured evidence based on the candidate argument text fragment set using functional type labeling, role constraint mapping, and evidence item encapsulation mechanisms, and outputting a structured evidence set, specifically includes: Step S301: Assign at least one broad argument function type to each candidate argument text fragment in the candidate argument text fragment set, the broad argument function type including at least one of participant, location, time, tool, affected entity, content, value, result and others; Step S302: Read the allowed role list from the event extraction index instance, perform compatible role mapping on the candidate argument text fragments with assigned broad argument function types according to the allowed role list, and output the compatible role set corresponding to each candidate argument text fragment; Step S303: Record auxiliary metadata for each candidate argument text fragment, wherein the auxiliary metadata includes at least one of evidence source, confidence level, lexical distance to the event trigger word, and matching support information; Step S304: Encapsulate each candidate argument text fragment, the lexical offset of the candidate argument text fragment, the broad argument function type, the compatible role set, and the auxiliary metadata into an evidence item, and denote the input context as... Let the list of allowed roles be denoted as The evidence item is represented as follows: in, Indicates the first One piece of evidence, Indicates the first A candidate argument text fragment, Indicates the first The initial lexical offset of each candidate argument text fragment. Indicates the first End-word offset of candidate argument text fragments Indicates the first A set of broad argument function types corresponding to each candidate argument text fragment. Indicates the first A set of compatible roles corresponding to each candidate argument text fragment, and satisfying the following conditions: , Indicates the first Auxiliary metadata corresponding to each candidate argument text fragment; Step S305: The structured evidence set is composed of multiple evidence items.

5. The evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in claim 4, characterized in that, Step S40, which involves generating argument / role pairs based on the structured evidence set using an evidence-guided prompting construction and JSON format constraint generation mechanism, and outputting intermediate structured generation results, specifically includes: Step S401: Merge identical text fragments and normalize identical text fragments in the evidence items of the structured evidence set, and output a deduplicated evidence set; Step S402: Extract the evidence item corresponding to each allowed role from the deduplicated evidence set according to the preset number of role evidence to retain, and construct a compact evidence table; Step S403: Extract the top-ranked evidence items from the deduplicated evidence set according to the preset global evidence retention quantity, and construct a global evidence list; Step S404: Write the event extraction index instance, the compact evidence table, the global evidence list, and the preset output format constraints into the evidence guiding argument generation prompt words; Step S405: Input the evidence-guided argument generation prompt into the large language model that follows the instructions, so that the large language model generates candidate argument role records according to the preset output format constraints, and outputs the intermediate structured generation result; wherein, the candidate argument role record includes an argument text field, a role field, and an evidence usage field, and the compact evidence table and the global evidence list are only used as original text location guidance information, and are not used as a fixed candidate set to limit the generation range of the large language model.

6. The evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in claim 5, characterized in that, Step S50, which involves performing location verification and correction based on the intermediate structured generation result using mechanisms such as original text fragment alignment, role validity verification, duplicate record deletion, and fragment boundary correction, and outputting the event argument extraction result, specifically includes: Step S501: Read each candidate argument role record from the intermediate structured generation result, perform exact string matching on the argument text field in the candidate argument role record in the input context, and output the first original text aligned record; Step S502: When the first original text alignment record is not obtained, perform case folding, punctuation normalization and whitespace character normalization on the argument text field, and perform string matching between the normalized argument text field and the input context to output the second original text alignment record; Step S503: When the first original text alignment record or the second original text alignment record includes multiple candidate alignment positions, the candidate alignment position with evidence item support is selected first; when multiple candidate alignment positions are supported by evidence items, the candidate alignment position with the smallest lexical distance to the event trigger word index is selected. Step S504: Compare the role field in the candidate argument role record with the allowed role list, delete candidate argument role records whose role field is not included in the allowed role list, and output a set of valid role records; Step S505: Deduplicate candidate argument role records in the legal role record set that have the same role and the same or overlapping text fragments, and retain candidate argument role records that meet the preset retention conditions in terms of original text alignment and evidence support. Step S506: Perform fragment boundary correction on the candidate argument role records after deduplication, and output the event argument extraction result. The event argument extraction result includes argument text fragments that can be located in the input context and their corresponding legal roles.

7. The evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in claim 6, characterized in that, Step S506, which involves performing fragment boundary correction on the candidate argument role records after deduplication, specifically includes: Read the aligned text fragments and role fields from the candidate argument role records after deduplication; Extract evidence items that are the same as or compatible with the role field from the structured evidence set, and output the evidence items to be compared; When the aligned text fragment satisfies the text fragment inclusion relationship with the candidate argument text fragment in the evidence item to be compared, the start word offset and end word offset of the aligned text fragment are replaced with the start word offset and end word offset of the candidate argument text fragment in the evidence item to be compared. When the aligned text fragment and the candidate argument text fragment in the evidence item to be compared do not satisfy the text fragment inclusion relationship, but the word overlap ratio of the two reaches the preset boundary correction threshold, the start word offset and end word offset of the aligned text fragment are replaced by the start word offset and end word offset of the candidate argument text fragment in the evidence item to be compared. When the preset recovery switch is enabled, text fragments that have been located in the input context, have a role in the allowed role list, and have evidence item support will be restored as candidate argument role records; The restored candidate argument role records are subjected to role validity checks and duplicate record deletion again, and the restored candidate argument role records are not generated by additional large language model generation calls.

8. An evidence-guided generation and original text location verification and correction system for zero-sample event argument extraction, characterized in that: The system includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the processor executes the computer program, it calls the following program units: The event instance formatting unit is used to acquire event text data to be processed, perform event instance formatting on the event text data to be processed using a lexical numbering and role list extraction mechanism, and output an event extraction index instance. The candidate fragment construction unit is used to construct candidate argument text fragments based on the event extraction index instance using original text fragment scanning and open evidence discovery prompting mechanism, and outputs a set of candidate argument text fragments. The structured evidence construction unit is used to perform structured evidence construction based on the candidate argument text fragment set using functional type labeling, role constraint mapping and evidence item encapsulation mechanisms, and output a structured evidence set; The evidence-guided generation unit is used to generate argument-role pairs based on the structured evidence set using evidence-guided prompts and a JSON format constraint generation mechanism, and output intermediate structured generation results. The original text positioning verification and correction unit is used to perform positioning verification and correction based on the intermediate structured generation result by means of original text fragment alignment, role legality verification, duplicate record deletion and fragment boundary correction mechanism, and output the event argument extraction result.

9. An evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction, characterized in that, The evidence-guided generation and original text location verification and correction device for zero-sample event argument extraction includes: a memory, a processor, and an evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction stored in the memory and executable on the processor. When the evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction is executed by the processor, it implements the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes an evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction. When the evidence-guided generation and original text location verification and correction program for zero-sample event argument extraction is executed by the processor, it implements the evidence-guided generation and original text location verification and correction method for zero-sample event argument extraction as described in any one of claims 1 to 7.

Images