Learning semantic and relationship joint embedding for author name disambiguation

Author name disambiguation is an important research topic in the academic information retrieval community. Existing methods rely either on feature engineering on rich attributes information or on relationship information to obtain documents’ similarity, but seldom consider the complementarity and the correlation between them. The feature engineering on attributes, especially on rich text information, could capture the global semantic concepts, while the relationship information could encode local structural proximity in multiple academic networks. To bridge the gap between semantic and relationship information in author name disambiguation, this paper presents a joint representation learning approach, which could encode both semantic and relationship information into a common low dimensional space. Specifically, the proposed method consists of four modules: (1) semantic embedding module; (2) relationship embedding module; (3) semantic and relationship joint embedding module; and (4) clustering module. Experimental results demonstrate that the proposed joint representation learning approach consistently outperforms the state-of-the-art methods on three benchmarks.