Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Triples-to-isiXhosa (T2X): Addressing the Challenges of Low-Resource Agglutinative Data-to-Text Generation

Francois Meyer, Jan Buys

TL;DR

This work targets data-to-text generation for low-resource, morphologically rich isiXhosa by introducing the T2X dataset and a dedicated evaluation framework. It proposes SSPG, a Subword Segmental Pointer Generator that jointly learns subword segmentation and copying, and shows its superiority among scratch models for isiXhosa and Finnish, while finding that fine-tuned bilingual MT models provide the strongest overall performance. The study reveals that standard pretrained language models underperform on this task, highlighting the limited transferability of off-the-shelf PLMs to low-resource agglutinative languages and the value of learning subword-aware generation. By extending experiments to Finnish Hockey and releasing both T2X and SSPG, the work offers a practical pathway for advancing data-to-text in underrepresented languages and underscores the need for language-specific subword techniques and MT-based pretraining strategies.

Abstract

Most data-to-text datasets are for English, so the difficulties of modelling data-to-text for low-resource languages are largely unexplored. In this paper we tackle data-to-text for isiXhosa, which is low-resource and agglutinative. We introduce Triples-to-isiXhosa (T2X), a new dataset based on a subset of WebNLG, which presents a new linguistic context that shifts modelling demands to subword-driven techniques. We also develop an evaluation framework for T2X that measures how accurately generated text describes the data. This enables future users of T2X to go beyond surface-level metrics in evaluation. On the modelling side we explore two classes of methods - dedicated data-to-text models trained from scratch and pretrained language models (PLMs). We propose a new dedicated architecture aimed at agglutinative data-to-text, the Subword Segmental Pointer Generator (SSPG). It jointly learns to segment words and copy entities, and outperforms existing dedicated models for 2 agglutinative languages (isiXhosa and Finnish). We investigate pretrained solutions for T2X, which reveals that standard PLMs come up short. Fine-tuning machine translation models emerges as the best method overall. These findings underscore the distinct challenge presented by T2X: neither well-established data-to-text architectures nor customary pretrained methodologies prove optimal. We conclude with a qualitative analysis of generation errors and an ablation study.

Triples-to-isiXhosa (T2X): Addressing the Challenges of Low-Resource Agglutinative Data-to-Text Generation

TL;DR

This work targets data-to-text generation for low-resource, morphologically rich isiXhosa by introducing the T2X dataset and a dedicated evaluation framework. It proposes SSPG, a Subword Segmental Pointer Generator that jointly learns subword segmentation and copying, and shows its superiority among scratch models for isiXhosa and Finnish, while finding that fine-tuned bilingual MT models provide the strongest overall performance. The study reveals that standard pretrained language models underperform on this task, highlighting the limited transferability of off-the-shelf PLMs to low-resource agglutinative languages and the value of learning subword-aware generation. By extending experiments to Finnish Hockey and releasing both T2X and SSPG, the work offers a practical pathway for advancing data-to-text in underrepresented languages and underscores the need for language-specific subword techniques and MT-based pretraining strategies.

Abstract

Most data-to-text datasets are for English, so the difficulties of modelling data-to-text for low-resource languages are largely unexplored. In this paper we tackle data-to-text for isiXhosa, which is low-resource and agglutinative. We introduce Triples-to-isiXhosa (T2X), a new dataset based on a subset of WebNLG, which presents a new linguistic context that shifts modelling demands to subword-driven techniques. We also develop an evaluation framework for T2X that measures how accurately generated text describes the data. This enables future users of T2X to go beyond surface-level metrics in evaluation. On the modelling side we explore two classes of methods - dedicated data-to-text models trained from scratch and pretrained language models (PLMs). We propose a new dedicated architecture aimed at agglutinative data-to-text, the Subword Segmental Pointer Generator (SSPG). It jointly learns to segment words and copy entities, and outperforms existing dedicated models for 2 agglutinative languages (isiXhosa and Finnish). We investigate pretrained solutions for T2X, which reveals that standard PLMs come up short. Fine-tuning machine translation models emerges as the best method overall. These findings underscore the distinct challenge presented by T2X: neither well-established data-to-text architectures nor customary pretrained methodologies prove optimal. We conclude with a qualitative analysis of generation errors and an ablation study.
Paper Structure (33 sections, 4 equations, 2 figures, 9 tables)

This paper contains 33 sections, 4 equations, 2 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Neural Data-to-text
  4. Subword Segmentation
  5. Triples-to-isiXhosa (T2X)
  6. Annotation
  7. Task difficulty
  8. Subword Segmental Pointer Generator (SSPG)
  9. BPE-based Data Encoder
  10. Subword Segmental Decoder
  11. Copying Segments
  12. Unmixed Decoding
  13. Experimental Setup
  14. Models
  15. Pointer generator (PG)
  16. ...and 18 more sections

Figures (2)

  • Figure 1: Example from T2X, showing the need for subword-based data-to-text modelling.
  • Figure 2: SSPG forward pass for (France, currency, Euro) $\rightarrow$ "Imali yaseFransi yi-Euro" ("The currency of France is the Euro"). At each character the next subword probability is computed with a mixture of a character-level decoder and a copy-equipped lexicon model (Eq. \ref{['mixture']}).