Unsupervised Pidgin Text Generation By Pivoting English Data and Self-Training

Ernie Chang, David Ifeoluwa Adelani, Xiaoyu Shen, Vera Demberg

Abstract

West African Pidgin English (WAPE) is spoken by over 75 million people but remains a low-resource language in Natural Language Processing. This paper presents an unsupervised approach to Natural Language Generation (NLG) for WAPE by leveraging English as a pivot language and employing self-training techniques. By bridging the domain gap between available parallel data (e.g., religious texts) and target domains (e.g., restaurant descriptions), the proposed method achieves a significant breakthrough in relevance while maintaining fluency.

Introduction

Despite its massive speaker base, WAPE suffers from a severe lack of high-quality parallel datasets required for training modern NLG systems. Existing machine translation models for Pidgin are typically trained on narrow datasets like the Bible or news articles. Consequently, these models exhibit poor performance when applied to different domains, such as describing structured restaurant data—a phenomenon known as domain mismatch. This research explores how to build a domain-specific Pidgin generator using only in-domain English data and unrelated monolingual Pidgin text.

Methodology

The authors propose a three-stage unsupervised pipeline to avoid the need for expensive human-labeled parallel data:

  1. Pivoting through English: A standard NLG system is trained to convert structured data (e.g., restaurant attributes) into English text, utilizing the high-resource nature of English.
  2. Unsupervised Neural Machine Translation (UNMT): A translation model is developed between in-domain English and out-domain Pidgin. Cross-lingual word embeddings (FastText) are used to align meanings across the two languages without direct parallel pairs.
  3. Self-Training: The model undergoes a bootstrapping phase where it generates "pseudo-parallel" pairs by translating English descriptions into Pidgin. These pairs are then used to refine the model, allowing it to adapt to the specific vocabulary and patterns of the target domain.

Evaluation Metrics

To assess the quality of the generated text, native Pidgin speakers evaluated outputs based on Relevance (0-1) and Fluency (0-2). The fluency scale provides a qualitative measure of how natural the generated Pidgin sounds to a native speaker, ranging from unintelligible to fluent.

Examples of Pidgin text fluency scoring used for human evaluation.

Results

The experimental results demonstrate the limitations of out-of-domain training and the efficacy of the self-training approach.

Quantitative comparison of model performance across Relevance and Fluency metrics.

As shown in the results table, the baseline model (PidginUNMT), trained on out-of-domain parallel data like the Bible, failed to capture the relevance of the restaurant domain, scoring a mere 0.038. In contrast, the self-trained model (model_self) achieved a relevance score of 0.434, a massive improvement. Notably, fluency remained stable across models, with the proposed approach achieving a score of 0.814, comparable to the baseline. Qualitative analysis showed that the model could produce coherent sentences such as "People dey feel the good food but all of us no dey available," which captures the essence of the data far better than previous attempts.

Discussion and Conclusion

This work highlights that domain alignment is critical for low-resource NLG. By using English as a scaffold and employing self-correction, the researchers successfully bridged the gap between general monolingual data and specific task requirements. While the system is not yet ready for production—limited by a small monolingual corpus (56k sentences) and vocabulary gaps—it represents a significant step toward inclusive AI for West African Pidgin English. Future work will investigate hybrid statistical and neural approaches to better manage low-resource constraints.

Summary updated

Abstract

West African Pidgin English is a language that is significantly spoken in West Africa, consisting of at least 75 million speakers. Nevertheless, proper machine translation systems and relevant NLP datasets for pidgin English are virtually absent. In this work, we develop techniques targeted at bridging the gap between Pidgin English and English in the context of natural language generation. %As a proof of concept, we explore the proposed techniques in the area of data-to-text generation. By building upon the previously released monolingual Pidgin English text and parallel English data-to-text corpus, we hope to build a system that can automatically generate Pidgin English descriptions from structured data. We first train a data-to-English text generation system, before employing techniques in unsupervised neural machine translation and self-training to establish the Pidgin-to-English cross-lingual alignment. The human evaluation performed on the generated Pidgin texts shows that, though still far from being practically usable, the pivoting + self-training technique improves both Pidgin text fluency and relevance.
BibTeX
@inproceedings{chang2020unsupervised,
  author = {Ernie Chang and David Ifeoluwa Adelani and Xiaoyu Shen and Vera Demberg},
  title = {{Unsupervised Pidgin Text Generation By Pivoting English Data and Self-Training}},
  booktitle = {AfricaNLP},
  year = {2020},
  url = {https://openreview.net/forum?id=Lvq3yV-OmhJ},
}
Figure