Understanding when and why neural ranking models fail for an IR task via error analysis is an important part of the research cycle. Here we focus on the challenges of (ⅰ) identifying categories of difficult instances (a pair of question and response candidates) for which a neural ranker is ineffective and (ⅱ) improving neural ranking for such instances. To address both challenges we resort to slice-based learning (Chen et al., 2019) for which the goal is to improve effectiveness of neural models for slices (subsets) of data. We address challenge (ⅰ) by proposing different slicing functions (SFs) that select slices of the dataset-based on prior work we heuristically capture different failures of neural rankers. Then, for challenge (ⅱ) we adapt a neural ranking model to learn slice-aware representations, i.e. the adapted model learns to represent the question and responses differently based on the model's prediction of which slices they belong to. Our experimental results1 across three different ranking tasks and four corpora show that slice-based learning improves the effectiveness by an average of 2% over a neural ranker that is not slice-aware.
展开▼
