ListOPT: Learning to Optimize for XML Ranking

摘要

Many machine learning classification technologies such as boosting, support vector machine or neural networks have been applied to the ranking problem in information retrieval. However, since the purpose of these learning-to-rank methods is to directly acquire the sorted results based on the features of documents, they are unable to combine and utilize the existing ranking methods proven to be effective such as BM25 and PageRank. To solve this defect, we conducted a study on learning-to-optimize, which is to construct a learning model or method for optimizing the free parameters in ranking functions. This paper proposes a listwise learning-to-optimize process ListOPT and introduces three alternative differentiable query-level loss functions. The experimental results on the XML dataset of Wikipedia English show that these approaches can be successfully applied to tuning the parameters used in an existing highly cited ranking function BM25. Furthermore, we found that the formulas with optimized parameters indeed improve the effectiveness compared with the original ones.