Non-entity hotline appeal event argument extraction method based on boundary enhancement

Abstract

The invention discloses non-entity hotline appeal event argument extraction based on boundary enhancement, which specifically comprises the following steps of: firstly, constructing a model architecture taking BERT as a substrate, and enhancing context semantics and vocabulary features by adopting character-level embedding; secondly, a rotation position coding mechanism is innovatively introduced, a boundary sensing module fused with rotation position coding is designed, and the sensitivity of the model to an entity boundary is enhanced by utilizing explicit modeling of inter-word distance characteristics of a relative position matrix. And finally, aiming at a positive and negative sample and multi-class imbalance problem, designing a layered focus loss function, and effectively solving the class imbalance problem through a learning weight optimization argument classification task of a dynamic weight adjustment difficult sample and a low-frequency class. Experimental verification shows that compared with a traditional deep learning method, the method has remarkable boundary recognition precision improvement and overall performance advantages.

Description

Technical Field

[0001] This invention relates to event extraction methods in the field of natural language processing technology, specifically to a method for extracting non-entity hotline appeal event arguments based on boundary enhancement. Background Technology

[0002] Event argument extraction, as the most crucial step in event extraction, aims to identify text spans related to specific event types from unstructured text and assign them semantic role categories. As a key subtask of the event extraction system, its core challenge lies in accurately capturing the semantic boundaries and functional attributes of event participants. Existing research generally focuses on extracting entity arguments with clear boundaries, neglecting the crucial role of non-named entity arguments (such as temporal expressions, verb phrases, and complex semantic fragments). Non-entity arguments exhibit significant boundary ambiguity and context-dependent characteristics. For example, in citizen hotline complaints, text fragments such as "continuous water pipe leakage" and "severe water accumulation on the kitchen floor" carry core event elements, but their semantic spans are neither traditional named entities nor do they possess explicit boundary markers. Notably, long entity arguments of address types in citizen hotline complaints (such as "green belt on the southeast side of the intersection of XX Road and XX Street") also face boundary ambiguity issues, posing a serious challenge to traditional sequence labeling-based methods. Therefore, this study combines boundary enhancement techniques to address this problem. To improve the performance of event argument extraction, boundary enhancement is particularly important when dealing with non-entity event arguments. The boundary enhancement proposed in this invention enhances the model's ability to identify complex events by expanding and strengthening event boundaries in the original data. This method can effectively improve the extraction results of non-entity event arguments.

[0003] Traditional event argument extraction methods mainly revolve around rule-driven and statistical models. Early rule-based systems required customized manual feature templates for different domains, which had inherent defects such as weak generalization ability and high development cost. Subsequent paradigms based on statistical learning, such as Hidden Markov Model (HMM) (Yu Jiangde, Xiao Xinfeng, Fan Xiaozhong. Chinese text event information extraction based on Hidden Markov Model C] / / 2007 National Conference on Open Distributed and Parallel Computers.0[2025-04-18]) and Conditional Random Field (CRF) (Fu Jianfeng, Liu Zongtian, Liu Wei, Zhou Wen. Event causal relationship extraction based on cascaded conditional random fields [J]. Pattern Recognition and Artificial Intelligence, 2011, 24(4):567-573), although improved the model adaptability through sequence labeling mechanism, are still limited to the recognition of explicit entity arguments (such as named entities such as people's names, places, and organizations), and are difficult to effectively model the semantic features of unnamed event arguments. Traditional event argument extraction methods do not consider argument boundary information and have insufficient context modeling capabilities, which can lead to incorrect extraction or omission of unnamed event arguments.

[0004] The sequence labeling-based event argument extraction method (Xinyu Li, Fayuan Li, Lu Pan, Yuguang Chen, Weihua Peng, Quan Wang, Yajuan Lyu, and Yong Zhu. 2020. DuEE: A Large-Scale Dataset for Chinese Event Extraction in Real-World Scenarios. In Natural Language Processing and Chinese Computing: 9th CCF International Conference, NLPCC 2020, Zhengzhou, China, October 14–18, 2020, Proceedings, Part II. Springer-Verlag, Berlin, Heidelberg, 534–545. https: / / doi.org / 10.1007 / 978-3-030-60457-8_44) is proposed based on classification-based methods. This method uses labels to mark the start and end positions of each argument in a text sequence. For example, the BIOE label represents the start token of the event argument, the inner token of the event argument, the outer token of the event argument, and the end token of the event argument. Collobert et al. (Huang, Z., Xu, W., & Yu, K. (2015). Bidirectional LSTM-CRF Models for Sequence Tagging. arXiv preprint arXiv:1508.01991.) introduced linear chain conditional random fields into neural network-based sequence labeling models and used multilayer perceptrons and convolutional neural networks to avoid designing task-specific features for handling the labeling problem of various sequences.Huang et al. (Hiroki Ouchi, Jun Suzuki, Sosuke Kobayashi, Sho Yokoi, Tatsuki Kuribayashi, Ryuto Konno, and Kentaro Inui. 2020. Instance-Based Learning of Span Representations: A Case Study through Named Entity Recognition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6452–6459, Online. Association for Computational Linguistics.) introduced BiLSTM-CRF to construct a character-level representation model, thus transforming the event argument extraction problem into a sequence labeling problem. The above method is effective for extracting explicit entity arguments, but it cannot accurately extract non-named entity event arguments.

[0005] A span-based selection strategy provides a new approach to event argument extraction through a start-end position prediction mechanism. Ouchi et al. (Fei Li, ZhiChao Lin, Meishan Zhang, and Donghong Ji. 2021. A Span-Based Model for Joint Overlapped and Discontinuous Named Entity Recognition. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4814–4828, Online. Association for Computational Linguistics.) transformed the nested NER task into an instance-based learning task to learn the similarity between spans. They mapped the spans and event arguments in the training set to a vector space, and then calculated the optimal similarity between spans to obtain the span class. Li et al. (Yan, H., Deng, B., Li, X., & Qiu, X. (2020). TENER: Adapting Transformer Encoder for Named Entity Recognition. In Proceedings of the 28th International Conference on Computational Linguistics (pp. 7091–7100). International Committee on Computational Linguistics.) proposed another span-based NER model. This model identifies event argument fragments and classifies relationships by enumerating all possible spans. It uses an attention mechanism to guide a graph convolutional network, enhancing the network's performance and thus more effectively capturing relationships and contextual information between entities. However, existing methods often suffer from span truncation or semantic incompleteness due to the lack of explicit modeling of boundary features.

[0006] To address the aforementioned issues, this invention proposes RoEAE (Rope Event Argument Extraction), a boundary-enhanced method for extracting non-entity hotline complaint event arguments. This method incorporates a boundary-aware module that integrates Rotation Position Encoding (RoPE) and explicitly models inter-word distance features using a relative position matrix, enhancing the model's sensitivity to entity boundaries. To address the issues of positive / negative samples and multi-class imbalance, a hierarchical focus loss function is designed, dynamically adjusting the learning weights for difficult samples and low-frequency classes. Validation experiments were conducted on the DuEE_fin dataset and a hotline complaint domain dataset. The experiments demonstrate that this method exhibits significant advantages in dynamic boundary awareness and complex argument structure modeling, providing a new technical approach for event knowledge extraction.

[0007] In summary, the non-entity hotline appeal event argument extraction method based on boundary enhancement is an innovative research problem with significant research value. Summary of the Invention

[0008] This invention proposes a boundary-enhanced framework for extracting non-entity event arguments. It extracts events from unstructured raw event description text, resulting in structured event descriptions that include event type, event role, and event arguments. First, a BERT-based model architecture is constructed, employing character-level embeddings to enhance contextual semantics and lexical features. Second, a rotational position encoding mechanism is innovatively introduced, strengthening the spatial correlation between words by modeling the relative positional relationships of boundaries, thus enhancing the model's understanding of the relative relationships between the start and end positions of arguments. Finally, a focus loss function is used to optimize the argument classification task, effectively addressing the imbalance between non-argument and argument categories.

[0009] A boundary-enhanced method for extracting non-entity hotline complaint arguments specifically includes:

[0010] (1) Boundary Enhancement Module

[0011] First, the text in the text dataset is segmented into words. Then, the segmented results are vectorized using a pre-trained model to obtain word vectors. Next, a dual-channel feature projection mechanism provides an operable vector space for subsequent rotation position encoding. This module abandons the traditional sequence labeling framework and instead constructs a span selection model based on start-end dual-position prediction. This paradigm establishes an explicit boundary-aware mechanism by simultaneously predicting the start and end positions of arguments, effectively addressing the problem of insufficient modeling of implicit positional features in traditional methods.

[0012] (2) Layered focus loss module

[0013] In the hierarchical focus loss module, class imbalance is a common problem in event argument extraction tasks using a span selection strategy, particularly where the proportion of non-argument categories is much larger than that of other argument labels. To alleviate this issue, this invention uses hierarchical focus loss instead of ordinary cross-entropy loss during training, making the model pay more attention to the classification of difficult argument roles. For inaccurately classified samples, the hierarchical focus loss remains unchanged; for accurately classified samples, the loss decreases.

[0014] The specific process of the method includes the following steps:

[0015] S1 first segments the original text of the input event into words, and then uses the BERT model to represent the segmented results into word vectors.

[0016] The S2 design employs a dual-channel feature projection mechanism to provide an operable vector space for subsequent rotational position encoding. This mechanism establishes independent feature spaces for different types of arguments while preserving the possibility of subsequent cross-type parameter sharing.

[0017] After obtaining the operable vector space, S3 introduces a dynamic boundary awareness factor in the position encoding layer to enhance the model's sensitivity to the features of the argument start and end positions. At the same time, a cross-layer position interaction mechanism is established, which explicitly models the relative distance relationship between the start and end positions through query-key position matrix operations in multi-head attention. Finally, the span representation method of SpanNER is integrated to jointly optimize the position encoding and span semantic representation, forming an end-to-end boundary reinforcement learning framework.

[0018] S4 calculates the score of an event argument for a single category by using the inner product of the vector spaces of the start and end positions. Here, considering the combination of event arguments, the start position of the argument must appear before the end position of the argument. Therefore, we add a lower triangular matrix to exclude the case where the start and end positions are reversed.

[0019] In the S5 event argument extraction task using a span selection strategy, class imbalance is a common problem, especially since the proportion of non-argument roles in the classification is much larger than that of other argument labels. To alleviate this problem, we use hierarchical focus loss instead of ordinary cross-entropy loss during training, making the model pay more attention to the classification of difficult argument roles. Attached Figure Description

[0020] Figure 1 Here is the algorithm structure diagram for the boundary-enhanced non-physical hotline appeal event argument extraction method.

[0021] Figure 2 Comparison of experimental results for complaint and reporting datasets

[0022] Figure 3Comparative experimental results for the DuEE Chinese text classification dataset

[0023] Figure 4 Setting up ablation experiments

[0024] Figure 5 Ablation experimental results for each module

[0025] Figure 6 Input-output graph for event extraction Detailed Implementation

[0026] To make the technical solution of the present invention clearer, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0027] A method for extracting non-entity event arguments from hotline requests based on boundary enhancement specifically includes:

[0028] (1) Boundary Enhancement Module

[0029] First, the text dataset The text x in i Word segmentation is performed to obtain Where w j For x i The result after word segmentation, where k is the number of words after segmentation, and d is the result of x. i The dimension of a vector The vector space is represented. The segmented results are then vectorized using a pre-trained model to obtain word vectors.

[0030] We compute Q,K = Split(Dense(E)), where Dense represents expanding the hidden layer dimension to 2d, and generate the query vector Q and key vector K respectively through the tensor splitting operation Split. Here, we transform Q and K into a start vector q representing the starting position using a rotation matrix. i and the end vector k representing the end position j Each start vector is treated as a query, and each end vector is retrieved to obtain a score. The span [i:j] is used as the score for category t, and the score is calculated. Obtain the t-class score of the event argument.

[0031] (2) Layered focus loss module

[0032] Class imbalance is a common problem in span selection strategies, especially when the proportion of non-argument categories is much larger than that of other argument labels. To alleviate this problem, this invention uses hierarchical focus loss instead of ordinary cross-entropy loss during training. Where, α tγ is the class weight, a trainable parameter, and γ is the hard sample focusing parameter, making the model pay more attention to the classification of hard argument roles. The hierarchical focus loss has no change for inaccurately classified samples, but the loss decreases for accurately classified samples.

[0033] The specific process of the method includes the following steps:

[0034] S1 This invention relates to the field of Natural Language Processing (NLP), specifically to a text representation method based on a pre-trained language model, applicable to various language understanding tasks, including event argument extraction. In one embodiment of this invention, the input original text sequence is first represented as a token sequence of length n, denoted as:

[0035] S = [w1, w2, ..., w n ]

[0036] Among them, w i This represents the i-th language unit that has undergone word segmentation. To achieve structured text processing, the word segmentation process employs the WordPiece segmentation method widely used in the BERT model. This method can divide words into sub-word units, effectively alleviating the out-of-vocabulary (OV) word problem and enhancing the model's ability to express the internal structure of words, thereby improving the accuracy of semantic modeling.

[0037] After word segmentation, the token sequence is input into the pre-trained language model BERT, where its embedding layer obtains the corresponding word vector representations. Specifically, the embedding representation of the text sequence can be denoted as:

[0038] E = BERT(S) = [e1, e2, ..., e n ]

[0039] Among them, e i This represents the d-dimensional word embedding vector corresponding to the i-th token. For a batch of input data, the output vector is represented as a three-dimensional tensor. L represents the length of the input sequence, and d is the hidden layer dimension of the model.

[0040] Through the above representation, the BERT model can generate word-level feature vectors with contextual semantic awareness, thus providing a high-quality semantic feature foundation for subsequent event argument extraction.

[0041] Subsequently, a dual-channel feature projection mechanism is designed to provide an operable vector space for subsequent rotation position encoding. Feature decoupling is achieved through a learnable dense transform layer.

[0042] Q,K = Split(Dense(E))

[0043] The Dense layer expands the hidden layer dimension to 2D and generates query vectors separately through a tensor splitting operation (Split). and key vector T represents the number of predefined event argument types. This design establishes independent feature spaces for different types of arguments, while preserving the possibility of subsequent cross-type parameter sharing.

[0044] This projection process employs a dimensionality reorganization strategy to decouple features. First, the feature tensor is expanded along the type dimension, and then a dual-path projection is constructed through channel segmentation. This approach ensures computational efficiency while enabling differentiated feature learning for different argument types through type-specific parameter matrices, laying the foundation for fine-grained feature representation in subsequent boundary enhancement layers.

[0045] S2 This paper proposes a span selection method based on boundary awareness enhancement for the event argument extraction task. Compared with the problems of strong label dependency and ambiguous boundary recognition in traditional sequence labeling paradigms, this method significantly improves the recognition accuracy of argument boundaries by establishing a dual-pointer span selection mechanism and combining it with improved relative position encoding.

[0046] At the architectural design level, we abandoned the traditional sequence labeling framework and instead constructed a span selection model based on start-end dual-position prediction. This paradigm establishes an explicit boundary-aware mechanism by simultaneously predicting the start and end positions of arguments, effectively addressing the problem of insufficient modeling of implicit positional features in traditional methods. This design not only conforms to the basic characteristics of event arguments as continuous text segments but also strengthens the coupling relationship of boundary features through interaction modeling between endpoints.

[0047] The boundary awareness mechanism is decomposed into 2 dimensions. In the attention mechanism, positional information is added when calculating the q and k vectors of the word embedding vector. The function formula is expressed as follows:

[0048] q m =f q (x m ,m)

[0049] k n =f k (x n ,n)

[0050] Where, q m Let x represent the word vector corresponding to the m-th token. m The query vector after integrating location information m, and k n Then it represents the word vector x corresponding to the nth token. nThe key vector after integrating location information n.

[0051] Span-based selection strategies primarily focus on constructing a suitable f(q,k) function. The calculation of the m-th word embedding vector x... m The corresponding self-attention output is q. m Other k n Calculate an attention score O for each. m :

[0052]

[0053]

[0054] Where d is the dimension of the hidden layer, k n Then it represents the word vector x corresponding to the nth token. n The key vector after integrating location information n.

[0055] Inspired by the above, this study makes targeted improvements to Rotation Position Encoding (RoPE) to enhance the model's ability to model relative positional relationships in text sequences. Compared to the insufficient length generalization of the absolute position encoding method used in the traditional Transformer architecture, RoPE can effectively model long-distance dependencies and maintain the relative position invariance of the sequence by embedding positional information into the rotation matrix transformation. Specifically, we make the following optimizations to the original RoPE: (1) Introduce a dynamic boundary awareness factor into the position encoding layer to enhance the model's sensitivity to the features of the argument start / end position; (2) Establish a cross-layer positional interaction mechanism to explicitly model the relative distance relationship between the start and end positions through query-key position matrix operations in multi-head attention; (3) Integrate the span representation method of SpanNER to jointly optimize the position encoding and span semantic representation to form an end-to-end boundary reinforcement learning framework.

[0056] This improved rotational positional encoding mechanism has dual advantages: at the feature modeling level, it maintains the translation invariance of relative positional relationships through rotation matrix transformation, effectively alleviating the positional encoding mismatch problem under different text lengths; at the task adaptation level, the unique boundary-aware factor design enables the model to accurately capture the contextual feature patterns of argument boundaries, and together with the span selection mechanism, it forms a collaborative optimization path of "position awareness-semantic understanding-boundary determination".

[0057] For Q and K, representing the token as the argument start and argument end vectors respectively, this paper uses RoPE positional encoding to perform rotation mapping:

[0058] qi =R i Q

[0059] k j =R j K

[0060] Where i represents the starting position of the event argument, j represents the ending position of the event argument, and q i and k j This represents the result after rotation operations on Q and K, with the rotation transformation matrix being R. θ :

[0061]

[0062] By rotation transformation, q i and k j It possesses direction-awareness in semantic space, thereby enhancing the ability to model argument boundaries. Using the transformed q... i and k j When you apply Attention to a sequence, the Attention mechanism automatically includes information about the relative positional boundaries.

[0063] RoPE enhances the model's ability to understand the relative relationships between the start and end positions of arguments, thus making it more conducive to accurately modeling the interactions of argument boundaries in event extraction;

[0064] S3 above uses a rotation matrix to transform Q and K into a start vector q representing the initial position. i and the end vector k representing the end position j Each start vector is treated as a query, and each end vector is queried to obtain a score, which is used as the argument category score.

[0065] Here, taking category t as an example, the boundary enhancement layer ultimately calculates the probability that the token is the start or end of an argument, and uses the inner product calculation method to obtain the score for the argument of category t:

[0066]

[0067] Here, considering the combination of event arguments, the starting position of an event argument must appear before the ending position of the event argument. Therefore, we add a lower triangular matrix. Used to exclude cases where the start and end positions are reversed.

[0068]

[0069] For different argument categories t, we have obtained the score s for each span[i;j]. The problem then becomes a multi-label classification problem of selecting k true arguments from n(n+1) / 2 candidate texts:

[0070]

[0071] In the event argument extraction task using the span selection strategy in S4, class imbalance is a common problem, especially since the proportion of non-argument classes is much larger than that of other argument labels. To alleviate this issue, we use hierarchical focus loss instead of ordinary cross-entropy loss during training, making the model pay more attention to the classification of difficult argument roles. For inaccurately classified samples, the hierarchical focus loss does not change the loss; for accurately classified samples, the loss decreases. Overall, it is equivalent to increasing the weight of inaccurately classified samples in the loss function. The hierarchical focus loss introduces an adjustment factor to calculate the loss, and the final loss function is as follows:

[0072]

[0073] Where, α t γ represents the class weights, and γ is the hard sample focusing parameter.

[0074] S5 This invention uses two datasets to verify the effectiveness of the proposed method: a complaint and reporting dataset and the Chinese event extraction dataset DuEE-Fin. The complaint and reporting dataset includes six role categories: request type, request time, request address, request subject, request phenomenon, and request assertion. The DuEE-Fin dataset contains 11,699 samples, with 7,015, 1,171, and 3,513 passages used for training, validation, and testing, respectively. In the event argument extraction task, both datasets are divided into training, validation, and testing sets in an 8:1:1 ratio.

[0075] like Figure 2 and Figure 3As shown, to verify the effectiveness of the boundary-enhanced event argument extraction method, the performance of different event argument extraction methods on the complaint and reporting dataset and the Chinese public dataset DuEE-fin was compared. On the hotline complaint dataset, the proposed model achieved an F1 score of 76.42%, significantly outperforming SpanNER (74.56%) and BOPN (74.67%), with improvements of 1.86% and 1.75%, respectively. This dataset involves complex event types and diverse language expressions, indicating that the proposed method has a strong ability to combat textual ambiguity and syntactic variability. On the Chinese public dataset DuEE-fin, the proposed method also outperforms TENER (50.03%), Bert-BiLSTM-CRF (69.03%), NFLAT (65.03%), and SpanNER (71.71%), achieving an F1 score of 73.12%, further verifying the model's adaptability to complaint texts with diverse structures and ambiguous entity boundaries. It should be noted that BOPN could not obtain experimental results on this dataset and could not make direct comparisons, but the performance of this invention in this scenario has surpassed most strong baseline models.

[0076] like Figure 4 and Figure 5 As shown, to verify the effectiveness of the boundary enhancement module and the hierarchical focus loss module in event argument extraction, ablation experiments were conducted on the hotline complaint dataset using the Spanner method with a span selection strategy as a benchmark, comparing EnRope and R_loss. E1 represents the experiment using the benchmark model Spanner, extracting event arguments based on a single span selection strategy; E2 represents adding a boundary enhancement module to enhance the boundary awareness of event arguments before extracting event arguments; E3 represents the method proposed in this invention, which combines the span selection strategy, the boundary enhancement module, and the hierarchical focus loss module. A comparison of E1 and E2 shows that interactive modeling of the start and end positions of arguments effectively improves the model's ability to recognize entity boundaries; a comparison of E2 and E3 shows the superiority of the improved loss function classification method compared to the cross-entropy loss method, especially in complex scenarios, improving recall and significantly optimizing overall extraction performance.

[0077] The model proposed in this invention not only improves the overall accuracy of event argument extraction compared to traditional methods, but also demonstrates good robustness and generalization in challenging tasks such as long-distance dependencies and boundary ambiguity. This is attributed to the model's multi-layered design in semantic modeling and structure awareness, which effectively captures the deep relationships between event elements and textual context, thereby improving extraction accuracy. Significant improvements were achieved on the hotline complaint dataset and the Chinese event extraction dataset DuEE-Fin, and the effectiveness of the boundary enhancement module and hierarchical focus loss module was verified through ablation experiments.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting non-entity event arguments from hotline requests based on boundary enhancement, characterized in that... include: (1) Boundary Enhancement Module First, the text in the text dataset is segmented into words. Then, the segmented results are vectorized into word vectors through a pre-trained model. After that, a dual-channel feature projection mechanism is used to provide an operable vector space for subsequent rotation position encoding. (2) Layered focus loss module During training, hierarchical focus loss is used instead of ordinary cross-entropy loss, making the model pay more attention to the classification of difficult-to-distinguish argument roles; The specific process includes the following steps: S1 first segments the original text of the input event into words, and then uses the BERT model to represent the segmented results into word vectors. The S2 design employs a dual-channel feature projection mechanism to provide an operable vector space for subsequent rotational position encoding. This mechanism establishes independent feature spaces for different types of arguments while preserving the possibility of subsequent cross-type parameter sharing. After obtaining the operable vector space, S3 introduces a dynamic boundary awareness factor in the position encoding layer to enhance the model's sensitivity to the features of the argument start and end positions. At the same time, a cross-layer position interaction mechanism is established, which explicitly models the relative distance relationship between the start and end positions through query-key position matrix operations in multi-head attention. Finally, the span representation method of SpanNER is integrated to jointly optimize the position encoding and span semantic representation, forming an end-to-end boundary reinforcement learning framework. S4 obtains the score of an event argument for a single category by calculating the inner product of the vector spaces of the starting and ending positions. Considering the combination of event arguments, the starting position of the argument must appear before the ending position of the argument. A lower triangular matrix is ​​added at the starting position of the argument to exclude the case where the starting and ending positions are reversed. In the S5 event argument extraction task, which employs a span selection strategy, hierarchical focus loss is used instead of ordinary cross-entropy loss during training.

2. The method for extracting non-entity event arguments from hotline requests based on boundary enhancement as described in claim 1, characterized in that... Specifically, it includes: (1) Boundary Enhancement Module First, the text dataset The text x in i Word segmentation is performed to obtain Where w j For x i The result after word segmentation, where k is the number of words after segmentation, and d is the result of x. i The dimension of a vector Represents a vector space; The segmented results are then vectorized using a pre-trained model to obtain word vectors. The query vector Q and the key vector K are generated by calculating Q,K = Split(Dense(E)), where Dense represents expanding the hidden layer dimension to 2d, and the query vector Q and the key vector K are generated by the tensor splitting operation Split. Here, Q and K are transformed into a start vector q representing the starting position by a rotation matrix. i and the end vector k representing the end position j Each word vector at the beginning position is used as a query vector, and the score is obtained by querying the word vector at the end position. Calculate the span [i:j] as the score for category t, and calculate... Obtain the t-class score of the event argument; (2) Layered focus loss module Hierarchical focus loss is used instead of ordinary cross-entropy loss during training. Where, α t γ is the class weight, a trainable parameter, and γ is the hard sample focusing parameter, which makes the model pay more attention to the classification of hard argument roles. The specific process includes the following steps: The original text sequence input by S1 is first represented as a token sequence of length n, denoted as: S=[w1,w2,…,w n ] Among them, w i This represents the i-th language unit that has undergone word segmentation. To achieve structured text processing, the word segmentation process uses the WordPiece segmentation method, which is widely used in the BERT model. After word segmentation, the token sequence will be input into the pre-trained language model BERT, and the corresponding word vector representation will be obtained through its embedding layer; specifically, the embedding representation of the text sequence can be denoted as: E=BERT(S)=[e1,e2,…,e n ] Among them, e i This represents the d-dimensional word embedding vector corresponding to the i-th token. For a batch of input data, the output vector is represented as a three-dimensional tensor. Where B represents the batch size, L represents the length of the input sequence, and d is the hidden layer dimension of the model; Subsequently, a dual-channel feature projection mechanism is designed to provide an operable vector space for subsequent rotation position encoding; feature decoupling is achieved through a learnable dense transform layer. Q,K = Split(Dense(E)) The Dense layer expands the hidden layer dimension to 2D and generates query vectors separately through a tensor splitting operation (Split). and key vector T represents the number of predefined event argument types; S2 decomposes the boundary awareness mechanism into a 2-dimensional structure. In the attention mechanism, positional information is added while calculating the q and k vectors from the word embedding vectors. The function formula is expressed as follows: q m =f q (x m ,m) k n =f k (x n ,n) Where, q m Let x represent the word vector corresponding to the m-th token. m The query vector after integrating location information m, and k n Then it represents the word vector x corresponding to the nth token. n The key vector after integrating location information n; Span-based selection strategies primarily focus on constructing an f(q,k) function; while calculating the embedding vector x of the m-th word... m The corresponding self-attention output is q. m Other k n Calculate an attention score O for each. m : Where d is the dimension of the hidden layer, k n Then it represents the word vector x corresponding to the nth token. n The key vector after integrating location information n; The following optimizations were made to the original RoPE: (1) A dynamic boundary awareness factor was introduced into the position encoding layer to enhance the model’s sensitivity to the features of the argument start / end position; (2) A cross-layer position interaction mechanism was established, and the relative distance relationship between the start position and the end position was explicitly modeled through query-key position matrix operation in multi-head attention; (3) The span representation method of SpanNER was integrated to jointly optimize the position encoding and span semantic representation to form an end-to-end boundary reinforcement learning framework. Q and K represent the vectors of the token as the start and end positions of arguments, respectively, which are rotated and mapped using RoPE position encoding: q i =R j Q k j =R j K Where i represents the starting position of the event argument, j represents the ending position of the event argument; q i and k j This represents the result after rotation operations on Q and K, with the rotation transformation matrix being R. θ : By rotation transformation, q i and k j It possesses direction-awareness in semantic space, thereby enhancing the modeling ability of argument boundaries; using the transformed q i and k j When performing attention on a sequence, the attention mechanism automatically incorporates relative positional boundary information; Q and K are transformed into a start vector q representing the initial position. i and the end vector k representing the end position j ; S3 treats each start vector as a query and retrieves a score from each end vector, which is used as the argument category score. For category t, the boundary enhancement layer ultimately calculates the probability that the token is the start or end of an argument, using an inner product calculation method to obtain the score for an argument of category t: Considering the combination of event arguments, the starting position of an event argument must appear before the ending position of the event argument; therefore, a lower triangular matrix is ​​added here. Used to exclude cases where the start and end positions are reversed. Given the scores s for each span[i;j] for different argument categories t, the problem becomes a multi-label classification problem of selecting k true arguments from n(n+1) / 2 candidate texts: The S4 hierarchical focus loss incorporates an adjustment factor to calculate the loss, and the final loss function is as follows: Where, α t Here, γ represents the class weights, and γ is the hard sample focusing parameter. The model is optimized through a loss function to extract non-entity events.

Images