COAD: Contrastive Pre-training with Adversarial Fine-tuning for Zero-shot Expert Linking
Expert finding, a popular service provided by many online websites such as Expertise Finder, LinkedIn, and AMiner, benefits seeking consultants, collaborators, and candidate qualifications. However, its quality is suffered from a single source of support information for experts. This paper employs AMiner, a free online academic search and mining system, having collected more than over 100 million researcher profiles together with 200 million papers from multiple publication databases, as the basis for investigating the problem of expert linking, which aims at linking any external information of persons to experts in AMiner. A critical challenge is how to perform zero shot expert linking without any labeled linkages from the external information to AMiner experts, as it is infeasible to acquire sufficient labels for arbitrary external sources. Inspired by the success of self supervised learning in computer vision and natural language processing, we propose to train a self supervised expert linking model, which is first pretrained by contrastive learning on AMiner data to capture the common representation and matching patterns of experts across AMiner and external sources, and is then fine-tuned by adversarial learning on AMiner and the unlabeled external sources to improve the model transferability. Experimental results demonstrate that COAD significantly outperforms various baselines without contrastive learning of experts on two widely studied downstream tasks: author identification (improving up to 32.1 14.8 indicates the superiority of the proposed adversarial fine-tuning method compared with other domain adaptation ways (improving up to 2.3 HitRatio@1).
READ FULL TEXT