Abstract
Verifying and consolidating expert information across the web is a significant challenge due to data fragmentation across news, social media, and academic repositories. We present CODE, a novel framework for zero-shot expert linking. CODE leverages contrastive pre-training on a large-scale academic database (AMiner) to learn robust expert representations and employs adversarial fine-tuning to bridge the domain gap between structured databases and heterogeneous external sources like news articles or LinkedIn profiles. Crucially, CODE achieves high accuracy without requiring any labeled cross-domain training data, making it a scalable solution for real-world expert finding systems.
Introduction
Expert finding is a critical service for government agencies, corporate recruiters, and academic researchers seeking consultants or collaborators. However, comprehensive expert data is rarely found in a single location. While academic databases like AMiner or Google Scholar track publications, professional skills may be listed on LinkedIn, and current activities are often reported in news articles.
The core challenge lies in expert linking: identifying which external mentions (e.g., a name in a news snippet) correspond to which profile in a central database. Traditional entity linking methods rely on extensive labeled datasets mapping external sources to the target database. In the context of the open web, where sources are unpredictable and diverse, manual annotation is prohibitively expensive. This paper addresses the problem of zero-shot expert linking, aiming to learn a universal matching metric from a rich source domain that generalizes to unseen target domains without explicit cross-domain labels.
The CODE Framework
The authors propose the CODE framework, which consists of two distinct stages designed to handle the heterogeneity of expert data.
1. Contrastive Pre-training (CODE-pre)
The first stage focuses on learning expert representations using only the source database (AMiner). An "expert instance" is defined as a set of papers authored by an individual. Using a multi-lingual BERT encoder, the model processes the text of these papers.
A key innovation here is the move away from simple aggregation. Instead of averaging paper embeddings into a single vector—which often leads to "semantic drift" for multi-disciplinary experts—CODE utilizes an interaction-based metric function. It calculates fine-grained similarities between individual papers using an RBF kernel. The model is trained using a triplet loss, which ensures that instances belonging to the same expert are closer in the embedding space than instances from different experts.
2. Adversarial Fine-Tuning
To adapt the pre-trained model to external sources with different linguistic styles and topics, CODE employs adversarial domain adaptation. The architecture includes a shared generator for both domains and a private generator specifically for the external target domain. A domain discriminator, implemented with a Gradient Reversal Layer (GRL), forces the shared generator to produce features that are domain-invariant. This process essentially "strips away" the domain-specific noise of external sources (like news syntax), allowing the pre-trained matching metric to operate effectively on the transformed external data.
Key Insights: Interaction and Adaptation
The success of CODE stems from two primary design choices:
- Interaction Beats Aggregation: By comparing fine-grained interactions between individual papers or attributes rather than global averages, the model can better represent experts who work across multiple distinct fields.
- Asymmetric Adversarial Adaptation: Unlike standard domain adaptation that uses private encoders for both domains, CODE focuses solely on a private encoder for the external domain. Since the goal is to link to the academic database, the model prioritizes aligning external features with the established AMiner feature space.
Experimental Results
The framework was evaluated using three datasets: the AMiner (WhoIsWho) dataset for pre-training, and News and LinkedIn datasets for external linking evaluation.
Performance Metrics
- Intrinsic Evaluation: On AMiner author identification, CODE-pre achieved a HitRatio@1 (HR@1) of 0.898, significantly outperforming representation-based baselines like Camel and HetNetE by over 30%.
- Zero-Shot Linking: When applied to external sources without prior labels, the full CODE model achieved an HR@1 of 0.753 for News and 0.904 for LinkedIn.
- Ablation Study: Removing the adversarial domain discriminator led to a significant performance drop (~3% in HR@1), confirming its necessity for crossing the domain gap.
Real-World Deployment and Future Work
CODE has been successfully deployed in a live online system. To maintain and improve accuracy, the system incorporates an active learning loop where user feedback (such as "Thumbs Up" confirmations) is used to refine the model over time.
Despite its success, the authors note some limitations. The current model assumes the expert exists in the target database and does not yet handle "non-existing alignment" detection. Additionally, performance is currently constrained by GPU memory, which limits the number of negative samples in contrastive learning. Future work will focus on scaling the system to more diverse sources and addressing the "open-world" linking problem where some experts may not have pre-existing profiles.
Summary updated