A Case-Based Approach to Data-to-Text Generation

Abstract

Traditional Data-to-Text Generation (D2T) systems utilise carefully crafted domain specific rules and templates to generate high quality accurate texts. More recent approaches use neural systems to learn domain rules from the training data to produce very fluent and diverse texts. However, there is a trade-off with rule-based systems producing accurate text but that may lack variation, while learning-based systems produce more diverse texts but often with poorer accuracy. In this paper, we propose a Case-Based approach for D2T that mitigates the impact of this trade-off by dynamically selecting templates from the training corpora. In our approach we develop a novel case-alignment based, feature weighing method that is used to build an effective similarity measure. Extensive experimentation is performed on a sports domain dataset. Through Extractive Evaluation metrics, we demonstrate the benefit of the CBR system over a rule-based baseline and a neural benchmark.

A review of the development and future challenges of case-based reasoning

“…Yu et al (2021) applied the CBR method to the Text-to-SQL task in semantic parsing, following the classical CBR process, to implement the translation process from unstructured natural language to structured SQL. In contrast, Upadhyay et al (2021) proposed a CBR method to generate text from data. The method aims to balance the accuracy and diversity of the text summary generated by the D2T system.…”