Abstract

Reranking models have become ubiquitous in most NLP applications where suitable training data can be found, yet the utility of reranking has not been universal. This paper investigates two recent reranking “mysteries”, and finds that the use of features derived from baseline system scores is responsible for the underperformance in both cases. Several alternative methods for encoding features derived from baseline scores are shown to repair these problems. Additional methods for deriving score features from the output of baseline systems are investigated, leading to a better understanding of the best practices for training reranking models under a variety of conditions. 1

Citations