Retrieval and Sorting of Scientific Documents Based on Stacked Embedding and Hybrid Attention Model
Binghao Zeng, Xuedong Tian · 2024
Making full use of mathematical formulas and their contextual information is crucial for enhancing the performance of scientific literature retrieval models, where mathematical formulas serve as core elements. The existing methods inadequately use formula structure and contextual information in situations involving mathematical formulas, and ignore the part-of-speech features contained in the context. A two stage scientific document retrieval method, based on stacked embedding and hybrid attention fusion part-of-speech features, was proposed in this paper. Initially, MathML in documents is used to learn the structural and semantic information of mathematical formulas, facilitating scientific document retrieval focused on mathematical expression. Subsequently, the document context is extracted, the context’s part-of-speech features are introduced into the model through stacked embeddings, and a hybrid attention mechanism is used to learn the dependency between part-of-speech and context, features are then generated to improve the rationality of retrieval result ranking. Experiments were performed on the NTCIR-12 dataset in which we expanded with Chinese literature. The mAP@10 is 0.865 and NDCG@10 is 0.863 respectively.