Improving Open-Domain Answer Sentence Selection by Distributed Clients with Privacy Preservation

Anonymous

Abstract: Open-domain answer sentence selection (OD-AS2), as a practical branch of open-domain question answering (OD-QA), aims to respond to a query by a potential answer sentence from a large-scale collection. A dense retrieval model plays a significant role across different solution paradigms, while its success depends heavily on sufficient labeled positive QA pairs and diverse hard negative sampling in contrastive learning. However, it is hard to satisfy such dependencies in a privacy-preserving distributed scenario, where in each client, less in-domain pairs and a relatively small collection cannot support effective dense retriever training. To alleviate this, we propose a brand-new learning framework for \textbf{P}rivacy-preserving \textbf{D}istributed O\textbf{D-AS2}, dubbed PDD-AS2. Built upon federated learning, it consists of a client-customized query encoding for better personalization and a cross-client negative sampling for learning effectiveness. To evaluate our learning framework, we first construct a new OD-AS2 dataset, called Fed-NewsQA, based on NewsQA to simulate distributed clients with different genre/domain data. Experiment results shows that our learning framework can outperform its baselines and exhibit its personalization ability.

Paper Type: long

Research Area: Information Retrieval and Text Mining