SEER : A Knapsack approach to Exemplar Selection for In-Context HybridQA

TL;DR

SEER introduces a knapsack-based method for selecting diverse and representative exemplars to improve in-context learning for hybrid question answering, effectively handling large contexts and diversity requirements.

Contribution

It formulates exemplar selection as a Knapsack ILP, enabling flexible incorporation of diversity and capacity constraints for hybrid QA tasks.

Findings

SEER outperforms previous exemplar selection methods on FinQA and TAT-QA.

The knapsack formulation effectively balances diversity and size constraints.

Improves reasoning accuracy in hybrid question answering.

Abstract

Question answering over hybrid contexts is a complex task, which requires the combination of information extracted from unstructured texts and structured tables in various ways. Recently, In-Context Learning demonstrated significant performance advances for reasoning tasks. In this paradigm, a large language model performs predictions based on a small set of supporting exemplars. The performance of In-Context Learning depends heavily on the selection procedure of the supporting exemplars, particularly in the case of HybridQA, where considering the diversity of reasoning chains and the large size of the hybrid contexts becomes crucial. In this work, we present Selection of ExEmplars for hybrid Reasoning (SEER), a novel method for selecting a set of exemplars that is both representative and diverse. The key novelty of SEER is that it formulates exemplar selection as a Knapsack Integer Linear Program. The Knapsack framework provides the flexibility to incorporate diversity constraints that prioritize exemplars with desirable attributes, and capacity constraints that ensure that the prompt size respects the provided capacity budgets. The effectiveness of SEER is demonstrated on FinQA and TAT-QA, two real-world benchmarks for HybridQA, where it outperforms previous exemplar selection methods.