T2X (Triples-to-isiXhosa)

T2X (Triples-to-isiXhosa) is a data-to-text dataset for isiXhosa. It was constructed by translating a subset of the English WebNLG dataset into isiXhosa: each example maps a (subject, relation, object) triple from DBPedia to one or more isiXhosa sentences describing it. It can be used to train and evaluate sequence-to-sequence models for generating isiXhosa text from structured data.

Example
Triple	(South Africa, leaderName, Cyril Ramaphosa)
English WebNLG	Cyril Ramaphosa is the leader of South Africa
T2X isiXhosa	uCyril Ramaphosa yinkokheli yoMzantsi Afrika

Example
Triple	(France, currency, Euro)
English WebNLG	The currency of France is the Euro
T2X isiXhosa	Imali yaseFransi yi-Euro

Dataset details

	Train	Valid	Test
English WebNLG 1-triples	3,114	392	388
T2X triples	2,413	391	378
T2X verbalisations	3,859	600	888

The data covers 15 DBPedia categories. Three categories (Astronaut, Athlete and WrittenWork) are not included in the training data, only in the validation and test sets. In the training and validation sets, only one isiXhosa verbalisation per triple is given for a majority of the 371 domains, while the test set has multiple verbalisations (up to 3) for most examples, corresponding to multiple possible phrasings.

Data annotation process

Annotators, first-language isiXhosa speakers who have studied the language at university level, were presented with the triples and English WebNLG verbalisations and asked to provide one or more isiXhosa translations that reflect the content of the triples while phrasing the translations as naturally as possible. Annotators discussed questions that arose during translation with each other, to ensure consistency across annotations.

Schema

{
  "id": "int64",
  "subject": "string",
  "relation": "string",
  "object": "string",
  "references": "list[string]"
}

Each row is one (subject, relation, object) triple paired with a list of one or more isiXhosa verbalisations (references). In train, a small number of triples are repeated across rows, once per available reference verbalisation.

Original file format

This repository also includes the dataset in its original release format under raw/, mirroring github.com/francois-meyer/t2x:

• {split}.data — one triple per line, e.g. __start_entity__ Aarhus Airport __end_entity__ __start_type__ cityServed __end_type__ __start_value__ Aarhus, Denmark __end_value__
• {split}.text — the matching isiXhosa verbalisation(s) per line, with multiple references separated by *#

A small number of encoding artifacts (mojibake from a UTF-8/Windows-1252 double-encoding issue, e.g. Suárez for Suárez) present in the original GitHub files have been corrected in both the structured and raw files here.

Usage

from datasets import load_dataset

dataset = load_dataset("uctnlp/t2x")
print(dataset)
print(dataset["train"][0])

License

This dataset is a derivative of WebNLG (licensed CC BY-NC-SA 4.0) and DBPedia content, and is released here under the same terms: CC BY-NC-SA 4.0.

Citation

If you use this dataset, please cite:

@inproceedings{meyer-buys-2024-triples,
    title = "Triples-to-isi{X}hosa ({T}2{X}): Addressing the Challenges of Low-Resource Agglutinative Data-to-Text Generation",
    author = "Meyer, Francois  and
      Buys, Jan",
    editor = "Calzolari, Nicoletta  and
      Kan, Min-Yen  and
      Hoste, Veronique  and
      Lenci, Alessandro  and
      Sakti, Sakriani  and
      Xue, Nianwen",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.1464/",
    pages = "16841--16854",
    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."
}