| Dialogue is a common communication scenario in daily life,and the dialogue text generated by dialogue contains a large amount of entity relationship information to be explored.Unlike formal texts with clear discourse structures,dialogue texts often contain pronouns and colloquial expressions,resulting in low effective information density and challenges for dialogue-based relation extraction.Recent studies have modeled dialogue-based relation extraction as a discourse-level task using graph neural networks and pre-trained language models,achieving good results.However,previous studies still face the following problems:(1)ignoring the structural and dependency information of the dialogue text itself and entities when modeling,focusing only on the characteristics of long texts;(2)neglecting the semantic connections between entities and their context in constructing entity features,as multiple entities and speakers can cause rapid topic switching;(3)assuming that relation types are equally independent during classification,ignoring the long-tail distribution and logical constraints in social relationships.To address these issues,this thesis proposes a Speaker-oriented Structural Dependency Model(SoSD)for dialogue relation extraction based on structural enhancement.The main contributions of the model are as follows:(1)To address the low information density of dialogue texts,a structure-enhanced attention encoding method is proposed.The structural information matrix centered on the speaker is generated based on the dialogue’s structural dependency relationships.During the encoding process using pre-trained language models,this structure information matrix is introduced as prior knowledge to modify the attention distribution,guiding the model to focus on relevant clues related to the speaker and ignore irrelevant information to obtain high-quality text features.(2)To address the problem of multiple entities and entity mentions in dialogues,dynamic entity mention features are constructed based on context information.Attention weights for feature fusion are calculated based on the semantic relevance between entity mentions and the current entity,and feature fusion is performed according to the weight distribution to construct high-quality global entity features.(3)To address the problem of complex relation types in dialogue relation extraction,a new loss function construction method is proposed.Adaptive threshold classes and Focal Loss are used to replace crossentropy as the classification loss to alleviate the long-tail distribution problem in label samples.The relation logic constraint loss is used to enhance the model’s ability to extract mutually inverse and symmetric relations.Experiments on public datasets demonstrate the effectiveness of the SoSD model.Each module is explained through ablation experiments and instance analysis.In addition,based on the algorithm,a visualization system for social network graphs in dialogue texts is constructed.Through complete requirements analysis and system design,a display system that supports model management,prediction,custom parameter training,and social network graph visualization functions is implemented. |
