Ìtàkúròso: Exploiting Cross-Lingual Transferability for Natural Language Generation of Dialogues in Low-Resource, African Languages

In this study, we investigate the possibility of cross-lingual transfer from a state-of-the-art (SotA) deep monolingual model DialoGPT to 6 African languages and compare with 2 other baselines (BlenderBot 90M, another SotA and a simple seq2seq). The languages are Swahili, Wolof, Hausa, Nigerian Pidgin English, Kinyarwanda & Yoruba. Natural language generation (NLG) of dialogues is known to be a challenging task for many reasons. It becomes more challenging for African languages which are low-resource in terms of data. We translate and train on a small portion of the multi-domain MultiWOZ dataset for the languages. Besides intrinsic evaluation (i.e. perplexity), we conduct human evaluation of single-turn conversations using majority voting and measure inter-annotator agreement (IAA) using Fleiss Kappa and credibility tests. The results show that the hypothesis that deep monolingual models learn some abstractions that generalise across languages holds. We observe human-like conversations in 5 out of the 6 languages. It, however, applies to different degrees in different languages, which is expected. The language with the most transferable properties is the Nigerian Pidgin English, with a human-likeness score of 78.1\%, of which 35.5\% are unanimous. Its credibility IAA unanimous score is 66.7\%. The main contributions of this paper include the representation of under-represented African languages and demonstrating the cross-lingual transferability hypothesis. We also provide the datasets and host the model checkpoints/demos on the HuggingFace hub for public access.