DOCENT: Learning Self-Supervised Entity Representations from Large Document Collections
This addresses the problem of entity representation learning for tasks like retrieval and knowledge base completion, offering a novel approach but with incremental improvements in specific domains.
The paper tackles learning self-supervised entity representations from large document collections by expanding context beyond local sentences to include all related text, enabling applications like retrieval and question answering. Results show models match or outperform baselines on tasks such as MovieLens tag prediction and natural language movie search, sometimes with minimal fine-tuning, and scale to large corpora like 1B-word Amazon reviews.
This paper explores learning rich self-supervised entity representations from large amounts of the associated text. Once pre-trained, these models become applicable to multiple entity-centric tasks such as ranked retrieval, knowledge base completion, question answering, and more. Unlike other methods that harvest self-supervision signals based merely on a local context within a sentence, we radically expand the notion of context to include any available text related to an entity. This enables a new class of powerful, high-capacity representations that can ultimately distill much of the useful information about an entity from multiple text sources, without any human supervision. We present several training strategies that, unlike prior approaches, learn to jointly predict words and entities -- strategies we compare experimentally on downstream tasks in the TV-Movies domain, such as MovieLens tag prediction from user reviews and natural language movie search. As evidenced by results, our models match or outperform competitive baselines, sometimes with little or no fine-tuning, and can scale to very large corpora. Finally, we make our datasets and pre-trained models publicly available. This includes Reviews2Movielens (see https://goo.gle/research-docent ), mapping the up to 1B word corpus of Amazon movie reviews (He and McAuley, 2016) to MovieLens tags (Harper and Konstan, 2016), as well as Reddit Movie Suggestions (see https://urikz.github.io/docent ) with natural language queries and corresponding community recommendations.