Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Curriculum Learning for Cross-Lingual Data-to-Text Generation With Noisy Data

Kancharla Aditya Hari, Manish Gupta, Vasudeva Varma

TL;DR

The paper tackles cross-lingual data-to-text generation with noisy data by applying curriculum learning. It introduces two training schedules (Expanding and Annealing) and a novel Alignment score that jointly accounts for input facts and target text, with the annealing strategy yielding the strongest gains. Experiments on XAlign and xToTTo across 11 Indian languages plus English show up to around 4 BLEU points and 5–15% improvements in faithfulness and coverage, validated by GPT-4 and human evaluators. The work demonstrates that data quality-aware curricula can outperform traditional criteria in multilingual noisy settings and provides code and data to support reproducibility.

Abstract

Curriculum learning has been used to improve the quality of text generation systems by ordering the training samples according to a particular schedule in various tasks. In the context of data-to-text generation (DTG), previous studies used various difficulty criteria to order the training samples for monolingual DTG. These criteria, however, do not generalize to the crosslingual variant of the problem and do not account for noisy data. We explore multiple criteria that can be used for improving the performance of cross-lingual DTG systems with noisy data using two curriculum schedules. Using the alignment score criterion for ordering samples and an annealing schedule to train the model, we show increase in BLEU score by up to 4 points, and improvements in faithfulness and coverage of generations by 5-15% on average across 11 Indian languages and English in 2 separate datasets. We make code and data publicly available

Curriculum Learning for Cross-Lingual Data-to-Text Generation With Noisy Data

TL;DR

The paper tackles cross-lingual data-to-text generation with noisy data by applying curriculum learning. It introduces two training schedules (Expanding and Annealing) and a novel Alignment score that jointly accounts for input facts and target text, with the annealing strategy yielding the strongest gains. Experiments on XAlign and xToTTo across 11 Indian languages plus English show up to around 4 BLEU points and 5–15% improvements in faithfulness and coverage, validated by GPT-4 and human evaluators. The work demonstrates that data quality-aware curricula can outperform traditional criteria in multilingual noisy settings and provides code and data to support reproducibility.

Abstract

Curriculum learning has been used to improve the quality of text generation systems by ordering the training samples according to a particular schedule in various tasks. In the context of data-to-text generation (DTG), previous studies used various difficulty criteria to order the training samples for monolingual DTG. These criteria, however, do not generalize to the crosslingual variant of the problem and do not account for noisy data. We explore multiple criteria that can be used for improving the performance of cross-lingual DTG systems with noisy data using two curriculum schedules. Using the alignment score criterion for ordering samples and an annealing schedule to train the model, we show increase in BLEU score by up to 4 points, and improvements in faithfulness and coverage of generations by 5-15% on average across 11 Indian languages and English in 2 separate datasets. We make code and data publicly available

Paper Structure

This paper contains 14 sections, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Curriculum Learning Strategy
  4. Experimental Setup
  5. Results and Analysis
  6. Conclusion
  7. Limitations
  8. Ethical Considerations
  9. Language Codes
  10. Hyper-parameters for Reproducibility
  11. LLM Prompts
  12. Alignment annotation prompt
  13. LLM-evaluation prompt
  14. Examples

Figures (4)

  • Figure 1: Examples of noisy data in the XAlign dataset 10.1145/3487553.3524265. The input facts are represented as RDF triples with head, relation and tail tags (<H>, <R> and <T>). Highlighted text in the reference cannot be inferred from the input facts.
  • Figure 2: Two curriculum schedules: (a) Expanding schedule introduces new shards, (b) annealing schedule removes shards as the training progresses. Each row represents a training phase.
  • Figure 5: Some examples of generation from XAlign using the best performing model compared to baseline model
  • Figure 6: Some examples of generation from xToTTo using the best performing model compared to baseline model