Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ARISE: Iterative Rule Induction and Synthetic Data Generation for Text Classification

Yashwanth M., Vaibhav Singh, Ayush Maheshwari, Amrith Krishna, Ganesh Ramakrishnan

TL;DR

ARISE presents a bootstrapping framework that jointly generates synthetic exemplars and induces generalized syntactic rules for text classification. By extracting higher-order syntactic-n-grams via Least General Generalization and filtering with a submodular graph-cut, ARISE creates a complementary supervision signal that can be used in in-context learning, fine-tuning, or joint learning. Across diverse full-shot, few-shot, and multilingual benchmarks, ARISE yields statistically significant gains and even SotA performance in several settings, demonstrating the value of combining structured linguistic cues with synthetic data. The approach scales through iterative refinement, showing practical impact for data-efficient text classification while highlighting trade-offs in computational overhead and the potential to leverage unlabeled data in future work.

Abstract

We propose ARISE, a framework that iteratively induces rules and generates synthetic data for text classification. We combine synthetic data generation and automatic rule induction, via bootstrapping, to iteratively filter the generated rules and data. We induce rules via inductive generalisation of syntactic n-grams, enabling us to capture a complementary source of supervision. These rules alone lead to performance gains in both, in-context learning (ICL) and fine-tuning (FT) settings. Similarly, use of augmented data from ARISE alone improves the performance for a model, outperforming configurations that rely on complex methods like contrastive learning. Further, our extensive experiments on various datasets covering three full-shot, eight few-shot and seven multilingual variant settings demonstrate that the rules and data we generate lead to performance improvements across these diverse domains and languages.

ARISE: Iterative Rule Induction and Synthetic Data Generation for Text Classification

TL;DR

ARISE presents a bootstrapping framework that jointly generates synthetic exemplars and induces generalized syntactic rules for text classification. By extracting higher-order syntactic-n-grams via Least General Generalization and filtering with a submodular graph-cut, ARISE creates a complementary supervision signal that can be used in in-context learning, fine-tuning, or joint learning. Across diverse full-shot, few-shot, and multilingual benchmarks, ARISE yields statistically significant gains and even SotA performance in several settings, demonstrating the value of combining structured linguistic cues with synthetic data. The approach scales through iterative refinement, showing practical impact for data-efficient text classification while highlighting trade-offs in computational overhead and the potential to leverage unlabeled data in future work.

Abstract

We propose ARISE, a framework that iteratively induces rules and generates synthetic data for text classification. We combine synthetic data generation and automatic rule induction, via bootstrapping, to iteratively filter the generated rules and data. We induce rules via inductive generalisation of syntactic n-grams, enabling us to capture a complementary source of supervision. These rules alone lead to performance gains in both, in-context learning (ICL) and fine-tuning (FT) settings. Similarly, use of augmented data from ARISE alone improves the performance for a model, outperforming configurations that rely on complex methods like contrastive learning. Further, our extensive experiments on various datasets covering three full-shot, eight few-shot and seven multilingual variant settings demonstrate that the rules and data we generate lead to performance improvements across these diverse domains and languages.

Paper Structure

This paper contains 30 sections, 8 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. ARISE - Automatic Rule Induction Using Syntactic Tree Generalization
  3. Rule Induction via LGG
  4. ARISE Framework
  5. Rule Filtering
  6. Bootstrapping Rules and Synthetic Data
  7. Seed and validation set during few-shot:
  8. Data Augmentation:
  9. Experiments
  10. Dataset:
  11. Data Generation:
  12. Experimental Setup
  13. In-context learning (ICL):
  14. Fine-tuning (FT):
  15. Joint Learning:
  16. ...and 15 more sections

Figures (2)

  • Figure 1: Overview of ARISE (Automatic Rule Induction using Syntactic tree gEneralization).
  • Figure 2: We induce rules via inductive generalization on syntactic n-grams, as shown (dependency relations omitted for brevety). The symbol '$\sqsupseteq$' denote a generalization operation. Trees labeled from $f_1$ to $f_5$ are instances of features. Similarly, trees labeled from $r_1$ to $r_8$ are rules.