Multi-perspective thought navigation for source-free entity linking

doi:10.1016/j.patrec.2023.12.020

Multi-perspective thought navigation for source-free entity linking

2024-02-01

0

PRE

AI

DOI：10.1016/j.patrec.2023.12.020

Original

Original request for help

Save

Abstract

En 中文

Neural entity-linking models excel at bridging the lexical gap of multiple facets of facts, such as entity-related claims or evidence documents. Despite advancements in self-supervised learning and pretrained language models, challenges persist in entity linking, particularly in interpretability and transferability. Moreover, these models need many aligned documents to adapt to emerging entities, which may not be available due to data scarcity. In this work, we propose a novel Demonstrative Self-TrAining fRamework (D-STAR) that leverages multi-perspective thought navigation. D-STAR iteratively optimizes a question generator and an entity retriever by navigating thoughts on a dynamic graph reasoning across multiple perspectives for question generation. The generated question-answer pairs, along with hard negatives shared in the graph, enable adaptation with minimal computational overhead. Additionally, we introduce a new task, source-free entity linking, focusing on unsupervised transfer learning without direct access to original domain data. To demonstrate the feasibility of this task, we provide a generated question-answering dataset, FandomWiki, for novel entities. Our experiments show that D-STAR significantly improves baselines on SciFact, Zeshel, and FandomWiki.

Keywords:

Information retrieval

Question generation

Entity linking

Chain-of-thought reasoning

Multi-perspective thought navigation for source-free entity linking

Abstract

Journal

Organization