Leveraging Contextual Information for Effective Entity Salience Detection
Salience detection with fine-tuned medium-sized language models.
Rajarshi Bhowmik, Marco Ponza, Atharva Tendle, Anant Gupta, Rebecca Jiang, Xingyu Lu, Qian Zhao, Daniel Preotiuc-Pietro
This paper investigates the efficacy of medium-sized language model with a cross-encoder style architecture in salience detection. Comprehensive benchmarking results demonstrate that the proposed architecture achieves significant improvement over previous feature engineering approaches.
Introduction
Salient entities are entities that are central to a piece of text, quanlified by either a binary or ordinal rating. Previous work explored heave feature engineering to craft explicit features to cover relevant aspects. This paper studies the effectiveness of Transformer-based Pre-trained Language Models (PLMs) for entity salience detection. The proposed method adopts a cross-encoder architecture where the entity or its alias along side its contextual mentions are encoded by a PLM. Then, a classifier uses the contextual representations and optionally positional information to determine the salience score of the target entity.
Related Work
Future Work
Utilizing external knowledge for saliance detection.