Parameter-Efficient Abstractive Question Answering over Tables or Text

TL;DR

This paper introduces a parameter-efficient approach for abstractive question answering over tables and text using encoder-decoder models with minimal additional parameters, outperforming or matching state-of-the-art results.

Contribution

It demonstrates that using only 0.7%-1.5% additional parameters with adapters can achieve competitive or superior QA performance across multiple datasets.

Findings

Outperforms state-of-the-art on tabular QA datasets

Achieves comparable results on textual QA with fewer parameters

Reduces training complexity while maintaining high performance

Abstract

A long-term ambition of information seeking QA systems is to reason over multi-modal contexts and generate natural answers to user queries. Today, memory intensive pre-trained language models are adapted to downstream tasks such as QA by fine-tuning the model on QA data in a specific modality like unstructured text or structured tables. To avoid training such memory-hungry models while utilizing a uniform architecture for each modality, parameter-efficient adapters add and train small task-specific bottle-neck layers between transformer layers. In this work, we study parameter-efficient abstractive QA in encoder-decoder models over structured tabular data and unstructured textual data using only 1.5% additional parameters for each modality. We also ablate over adapter layers in both encoder and decoder modules to study the efficiency-performance trade-off and demonstrate that reducing additional trainable parameters down to 0.7%-1.0% leads to comparable results. Our models out-perform current state-of-the-art models on tabular QA datasets such as Tablesum and FeTaQA, and achieve comparable performance on a textual QA dataset such as NarrativeQA using significantly less trainable parameters than fine-tuning.