Figure 1

The figure illustrates the architecture of our BAC-GNN-CRF NER model with an example. The input consists of an ancient Chinese sentence (\(\ldots , c_i,\ldots\)) and its corresponding chapter title (P) concatenated with SEP. The encoder layer generates contextualized embeddings (\(h_{ci}, h_p\)) for each character and the chapter title using the “Bert-Ancient-Chinese” model. In the GNN layer, the vertex set of the graph consists of the Chinese characters (\(c_i\)), matching words (\(w_i\)), and the chapter (P, serving as the global node). The words are extracted from a dictionary. The global node links all character nodes, while the word nodes link their corresponding character nodes. The CRF layer produces the final output tags for each character in the input sentence. Each tag consists of a prefix and its entity type (e.g. \(B_{OFI}\) denotes the beginning of an Office entity). The prefixes B-I-E-O stand for Begin, Inside, End, and Outside, respectively. The right part shows the process of incorporating matching words.