Speaker identification combining role knowledge graph correction and contextual block attention

Abstract

Character dialogue play a crucial role in novels, serving as a key element for understanding both the plot and character relationships. With advancements in artificial intelligence and natural language processing, dialogue speaker recognition has seen significant progress. In this paper, we integrate a character role knowledge graph with a self-training mechanism to perform the speaker recognition on Chinese novel quotations using large-scale novel corpora. Quotes in novels can generally be classified as explicit or implicit. Implicit quotes require identifying speakers from extensive contextual information, which poses challenges for existing models in handling long and detail-rich contexts. In addition, existing end-to-end speaker recognition methods are ineffective because they do not fully consider the relationship between context and quotes. In this paper, we first propose a Narrative Unit-based Context Selection (NUCS) algorithm for determining the context to which quotes belong. Secondly, a speaker recognition method based on Role Knowledge Graph Correction (RKGC) and Contextual Block Attention (CBA) is proposed. The proposed role knowledge graph correction algorithm improves speaker attribution by deeply analyzing role entities, their relationships, and relevant trigger words within quotations. It then refines candidate speaker probabilities obtained from the previous module. Additionally, the algorithm captures character relationships within the context using role mappings from existing novels. The CBA method introduces a block attention mechanism to capture character relationships within the context. It effectively computes the probability of each character being the speaker and determines the start and end indices of the most probable speaker segments. This allows the model to focus more precisely on speaker-related content, leading to more accurate speaker predictions. Experimental results demonstrate that our approach achieves competitive performance: on our self-constructed CNSI dataset, it attains 91.9% EM and 93.3% F1 scores. Although slightly lower than the SOTA SPC method (EM 92.3%, F1 93.6%), our approach demonstrates significant advantages in contextual reasoning capabilities, particularly for complex multi-character dialogue scenarios.