Cross-document event co-reference resolution and depolarization method based on anti-factual reasoning

Abstract

The invention belongs to the technical field of cross-document event co-reference resolution, and provides a cross-document event co-reference resolution depolarization method based on anti-fact reasoning, which comprises the steps of reference pair joint sequence construction, anti-fact perturbation, fact prediction probability modeling, deviation score modeling and dynamic correction. By dynamically subtracting a deviation component induced by a hybrid factor from total prediction, the model is forced to reduce dependence on vocabulary surface similarity, the model can be guided to turn to mining semantic clues hidden in a deep layer of a context when facing a complex negative sample, and the accuracy and robustness of the model are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cross-document event coreference resolution, in particular to a cross-document event coreference resolution debiasing method based on counterfactual reasoning. BACKGROUND

[0002] Event coreference resolution is abbreviated as ECR, and cross-document event coreference resolution is abbreviated as CD-ECR. ECR aims to identify different mentions in the text that point to the same event in the real world. CD-ECR needs to identify across different documents, and faces greater challenges. The existing technology usually regards CD-ECR as a similarity evaluation problem between event mentions, and uses a pre-trained language model to perform pair-wise representation learning.

[0003] However, the existing various technologies have many significant defects. First, the existing model often excessively relies on the lexical similarity of trigger words. If the trigger words of two event mentions are the same, the model tends to determine that they are coreferential, while ignoring the difference in context semantics. Conversely, the same is true. This judgment based on surface lexical matching leads to false correlation. In addition, due to the confusion of the relationship between lexical matching and real semantic coreference, the performance of the model will decrease significantly when dealing with negative samples with high lexical overlap but different semantics, or positive samples with different lexicons but the same semantics. Finally, the existing debiasing method often uses data augmentation, which usually requires a large amount of additional manual annotation or rule design, and the implementation cost is high and it is difficult to cover all cases. SUMMARY

[0004] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a cross-document event coreference resolution debiasing method based on counterfactual reasoning, which solves the defects of the existing method that excessively relies on the lexical similarity of trigger words, confuses the relationship between lexical matching and real semantic coreference, and requires a large amount of additional manual annotation or rule design.

[0005] To achieve the above purpose, the present application provides the following scheme: A cross-document event coreference resolution debiasing method based on counterfactual reasoning, comprising: structurally marking an input event mention, and embedding special placeholders at both ends of the trigger word of the event mention to obtain a joint sequence of mention pairs; counterfactually perturbing the joint sequence of mention pairs to obtain a counterfactual sequence; using a pre-set pair-wise event coreference resolution module to model the relationship of the joint sequence of mention pairs to obtain a fact prediction probability; using the pair-wise event coreference resolution module to model the relationship of the counterfactual sequence to obtain a bias score; dynamically correcting the fact prediction probability according to the bias score using a pre-set intervention formula to obtain a coreference decision.

[0006] Preferably, the referential pair joint sequence is: ; wherein, is the referential pair joint sequence; is a standard special classification token; is a standard special separation token; , is a text sequence of event referents , and a text sequence of event referents , respectively.

[0007] Preferably, the referential pair joint sequence is counterfactually perturbed to obtain a counterfactual sequence, comprising: masking contextual cues of non-trigger words in the referential pair joint sequence to obtain the counterfactual sequence.

[0008] Preferably, the paired event coreference resolution module comprises an input end, a cross encoder, a referential pair representation generation layer, an average pooling layer and a classifier connected in sequence.

[0009] Preferably, the intervention formula is: ; wherein, is the coreference decision; is the fact prediction probability; is a regulation factor; is the bias difference value.

[0010] Preferably, the expression of the referential pair representation generation layer is: ; wherein, is the generated referential pair representation; is a nonlinear transformation; is a weight matrix; is a global pooling representation of the input sequence; is a hidden layer embedding centered on the trigger word of the event referent ; is a hidden layer embedding centered on the trigger word of the event referent ; is an element-wise multiplication; denotes a matrix transposition operation.

[0011] The present application discloses the following technical effects: The present application provides a cross-document event coreference resolution debiasing method based on counterfactual reasoning, which dynamically subtracts the bias component induced by the mixed factor from the full prediction, solves the defects of the prior art such as excessive dependence on trigger word lexical similarity, confusion of lexical matching and real semantic coreference relationship, and the need for a large amount of additional manual annotation or rule design, and realizes the improvement of model accuracy and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor under the premise of the drawings.

[0013] Figure 1 A cross-document event coreference resolution debiasing process schematic diagram based on counterfactual reasoning is provided for the embodiments of the present application. Figure 2 A model overall framework diagram is provided for the embodiments of the present application. Figure 3 A pairwise-ECR architecture diagram is provided for the embodiments of the present application. Figure 4 An event coreference resolution structure causal diagram is provided for the embodiments of the present application. Figure 5 An event reference example in the intra-document and cross-document scenarios under the ECR task is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0015] The purpose of the present application is to provide a cross-document event coreference resolution debiasing method based on counterfactual reasoning, which solves the defects of excessive dependence on trigger word lexical similarity, confusion of lexical matching and real semantic coreference relationship, and the need for a large number of additional manual annotation or rule design in the prior art.

[0016] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0017] Figure 1 A cross-document event coreference resolution debiasing process schematic diagram based on counterfactual reasoning is provided for the embodiments of the present application, Figure 2 A model overall framework diagram is provided for the embodiments of the present application, as shown in Figure 1 and Figure 2 The present application provides a cross-document event coreference resolution debiasing method based on counterfactual reasoning, which comprises: Step 100: Perform structured tagging on the input event mentions and embed special placeholders at both ends of the trigger words of the event mentions to obtain a joint sequence of reference pairs; Step 200: Perform counterfactual perturbation on the joint sequence of the reference pairs to obtain a counterfactual sequence; Step 300: Use the preset pairwise event coreference resolution module to perform relation modeling on the joint sequence of the reference pairs to obtain the fact prediction probability; Step 400: Use the paired event coreference resolution module to perform relational modeling on the counterfactual sequence to obtain the bias score; Step 500: Based on the deviation score, dynamically correct the fact prediction probability using a preset intervention formula to obtain a co-reference decision.

[0018] Furthermore, the reference pair joint sequence is: ;in, For the joint sequence of the reference pairs; Standard special classification token; For standard special separator tokens; , Each refers to an event Text sequences and event references A text sequence.

[0019] Preferably, the counterfactual perturbation of the joint sequence of reference pairs yields a counterfactual sequence, including: The counterfactual sequence is obtained by masking the contextual clues of non-trigger words in the joint sequence of the reference pair.

[0020] refer to Figure 3 The paired event coreference resolution module includes: an input terminal, a cross encoder, a coreference pair representation generation layer, an average pooling layer, and a classifier connected in sequence.

[0021] Preferably, the intervention formula is: ;in, For the common-reference decision; Predict the probability of the stated fact; As a regulating factor; The deviation score is given.

[0022] Furthermore, the expression representing the generation layer is as follows: ;in, The generated reference pairs are represented; This is a non-linear transformation; This is the weight matrix; A global pooled representation of the input sequence; Referencing the event Hidden layer embedding centered on trigger words; Referencing the event Hidden layer embedding centered on trigger words; This is element-wise multiplication; This represents the matrix transpose operation.

[0023] refer to Figure 2 , Figure 3 , Figure 4 To address the issues of over-reliance on trigger word surface matching, spurious relevance, and neglect of deep contextual semantics in traditional pairwise classification models for cross-document event coreference resolution, this embodiment proposes abstracting the reasoning process into a structured causal model (SCM) to improve the accuracy and robustness of cross-document event coreference resolution. To achieve this, this embodiment defines three key variables: input variables... Represents the event mention and its context; output variables Represents the final coreference prediction result; variable This represents the trigger word matching feature as a confounding factor. By constructing this causal graph, this embodiment successfully identified spurious paths that interfered with coreference determination. It also reveals how trigger word matching masks the underlying causes through non-causal paths. arrive Based on the true semantic causal effect, a new paradigm for debiasing based on counterfactual reasoning is proposed to improve the accuracy and robustness of cross-document event coreference resolution tasks.

[0024] Specifically, this embodiment focuses on the identification and mechanism reconstruction of confounding effects based on a structured causal model. First, it deconstructs the causal logic of coreference discrimination. This embodiment overcomes the limitations of traditional black-box models that rely solely on statistical associations by introducing a structured causal model (SCM) to explicitly define the causal mechanism between event features and prediction results. Within this framework, three core variables are defined: independent variables representing event mentions and their full contextual information. The dependent variable representing the final core-referenced label prediction. And confounding factors characterizing the surface matching degree of trigger words. Through in-depth analysis of the causal graph topology, this embodiment identified a spurious path caused by confounding factors. (i.e., the "backdoor path"). This finding theoretically explains the root cause of the model's bias: the model often fails to fully exploit this bias during the judgment process. arrive The true semantic causal effect is not influenced by the lexical overlap of trigger words. The strong induction of this bias produces a spurious correlation driven by prejudice. This abstraction process provides solid theoretical support for subsequently mitigating the negative impact of confounding factors and extracting true causal effects.

[0025] Preferably, the system integrates multi-dimensional interactive factual reference representation extraction. After establishing the causal abstraction framework, the system enters the parallel processing stage of physical features, the core objective of which is to extract reference pair representations containing complete factual information. This scheme uses a cross-encoder based on the Transformer architecture as the backbone network. To significantly enhance the model's ability to perceive the core of events, the system performs structured tagging of event mentions at the input end by embedding special placeholders (such as...) at both ends of the trigger words. <m>and< / m> To highlight the core position. The reference pairs in the two documents are concatenated as follows: The combined sequence.

[0026] Furthermore, based on the hidden layer states generated by the encoder, this embodiment implements a feature aggregation strategy that integrates global context and local cores: utilizing the first-order... Vectors capture macroscopic semantics, while average-pooling is used to obtain local representations corresponding to trigger words. and To further explore the dynamic interactions between referential pairs, the system performs a non-linear mapping of global features, local features, and their element-wise products to construct a factual referential representation. The generated at this time As a high-dimensional vector, although it carries rich factual semantics, the discriminative cues are often deeply coupled with surface lexical features due to statistical dependencies in the model learning process. In order to effectively isolate bias-driven spurious contributions from this complex representation, the system must introduce a comparative benchmark that can quantify the intensity of bias.

[0027] Specifically, the deviation score is quantitatively mapped based on counterfactual perturbation. Based on the aforementioned need for deviation stripping, this embodiment designs a core counterfactual reasoning module, aiming to establish an opposing reference system for the aforementioned factual representations. This module follows the principles of stripping away true causality and isolating false correlations. It constructs an extreme hypothetical scenario where the system observes only the trigger word in the absence of all contextual information to measure the dependence of the prediction result on word overlap. Specifically, the system performs counterfactual perturbation on the original input: keeping the event trigger word unchanged, and completely masking all contextual cues of non-trigger words using a masking operation. This counterfactual input, containing only the lexical skeleton, is fed into a homogeneous encoder to generate the corresponding counterfactual reference representation. Subsequently, specialized classification weights are used to map these weights into digitized bias scores. This process achieves a digital mapping of lexical inducement bias, capturing the false confidence that the model generates when it lacks logical support. This digital mapping endows the model with deep introspection capabilities, enabling it to accurately define the interference weight of lexical overlap in the current sample on the final decision.

[0028] Preferably, the process involves causal debiasing inference and decision-making based on the total indirect effect. After quantitatively mapping the bias scores, the process enters the final debiasing inference stage. The core logic of this stage lies in: how to utilize the acquired bias scores... As a corrective instruction, non-causal interference is filtered out from the original fact predictions. This embodiment is based on the Total Indirect-Effect (TIE) principle and implements bias removal through causal intervention. Specifically, in the inference and judgment stage, the system no longer simply relies on the fact prediction probability containing lexical bias. Instead, it extracts the aforementioned steps. As a deduction item, through the intervention formula The judgment results are dynamically corrected. In this intervention mechanism... This subtraction operation balances debiasing intensity and semantic preservation. Through this subtraction, the system successfully eliminates confounding factors. The impact on output strongly guides the model to proactively delve into deeper semantic cues when faced with challenging samples. Significant performance improvements on the ECB+ and GVC datasets demonstrate the robustness and effectiveness of this logical chain from bias measurement to causal intervention. For specific examples of event referencing, please refer to [link to relevant documentation]. Figure 5 .

[0029] Specifically, Figure 2 , Figure 3The parameters are explained as follows: Factual Input and Counterfactual Input: The former contains the complete context and trigger words, while the latter only retains the trigger word features. Pairwise-ECR: The pairwise event coreference resolution module, responsible for modeling the relationship between two references. CrossEncoder: The core network that uses a Transformer structure to perform deep interaction modeling on the input sequence. Last hidden layer feature: The feature of the last hidden layer. Average pooling: The operation of compressing multiple word vectors into a fixed-length vector. Mention Pair Representation: A comprehensive vector that integrates global semantics and local trigger word features. Classifier and Coreference Decision: The classifier and coreference decision, the logical layer that ultimately determines whether two events refer to the same entity. Subtraction: Bias score and subtraction operation, representing the intervention process of deducting the contribution of lexical bias from fact prediction. Figure 4 The parameters are explained as follows: X (event mention pairs): Event mention pairs, representing the input reference samples to be evaluated. T (trigger word matching): Trigger word matching features, which act as confounders in this model. Due to their tendency to lead to spurious relevance, they are the core object of bias removal. Y (coreference labels): Coreference labels, representing the coreference classification results predicted by the model. Do-operator, Intervention: Intervention operations, which cut off the influence of confounders on independent variables through causal intervention, thereby obtaining the true causal effect.

[0030] The beneficial effects of this invention are as follows: This invention employs a causal intervention mechanism based on the total indirect effect. By dynamically subtracting bias scores from the prediction results of the full set of information, it forces the model to reduce its reliance on superficial lexical similarity. This logic compels the model to retrieve and utilize deep features such as context and event arguments to make decisions when faced with heterogeneous text, thereby improving the model's accuracy and robustness.

[0031] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0032] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for resolving cross-document event coreference and biasing based on counterfactual reasoning, characterized in that, include: The input event mentions are structured and marked, and special placeholders are embedded at both ends of the trigger words of the event mentions to obtain a joint sequence of reference pairs; The counterfactual perturbation of the joint sequence of claims yields a counterfactual sequence; The relationship modeling of the joint sequence of reference pairs is performed using a preset pairwise event coreference resolution module to obtain the fact prediction probability; The pairwise event coreference resolution module is used to model the relationship of the counterfactual sequence to obtain the bias score; Based on the deviation score, the predicted probability of the fact is dynamically corrected using a preset intervention formula to obtain a co-reference decision.

2. The method for resolving cross-document event coreference based on counterfactual reasoning and removing bias according to claim 1, characterized in that, The referencing pair of the joint sequence is: ;in, For the joint sequence of the reference pairs; Standard special classification token; For standard special separator tokens; , Each refers to an event Text sequences and event references A text sequence.

3. The method for resolving cross-document event coreference based on counterfactual reasoning and removing bias according to claim 1, characterized in that, The counterfactual perturbation of the joint sequence of the claims yields a counterfactual sequence, including: The counterfactual sequence is obtained by masking the contextual clues of non-trigger words in the joint sequence of the reference pair.

4. The method for resolving cross-document event coreference based on counterfactual reasoning as described in claim 1, characterized in that, The paired event coreference resolution module includes: an input terminal, a cross encoder, a coreference pair representation generation layer, an average pooling layer, and a classifier connected in sequence.

5. The method for resolving cross-document event coreference based on counterfactual reasoning as described in claim 1, characterized in that, The intervention formula is: ;in, For the common-reference decision; Predict the probability of the stated fact; As a regulating factor; The deviation score is given.

6. The method for resolving cross-document event coreference based on counterfactual reasoning according to claim 4, characterized in that, The expression representing the generation layer is: ;in, The generated reference pairs are represented; This is a non-linear transformation; This is the weight matrix; A global pooled representation of the input sequence; Referencing the event Hidden layer embedding centered on trigger words; Referencing the event Hidden layer embedding centered on trigger words; This is element-wise multiplication; This represents the matrix transpose operation.

Images