Abstract
Child language norming is essential for identifying speech and language delays. For South African languages such as Sesotho and Afrikaans, this process is severely bottlenecked by the slow pace of manual linguistic annotation. This project proposes to automate key components of the norming pipeline using deep learning models for linguistic annotation. Specifically, the project investigates the following tasks: Part-of-Speech (POS) tagging for Afrikaans and Sesotho child speech, and machine translation of Sesotho child utterances into English. Pre-trained multilingual language models, such as XLM-R (a massively multilingual model) and AfroXLM-R (an Africa-adapted variant), will be fine-tuned on adult-domain corpora before being evaluated on child speech samples from the CHILDES dataset. Evaluation will combine automatic metrics (accuracy, F1, BLEU) with a qualitative review by trained linguists. The result is a set of fine-tuned models and a comparative analysis that will inform future automated norming tools for under-resourced South African languages. This project assesses whether such models are accurate enough to support linguists as part of a human-in-the-loop annotation pipeline, in which automated predictions are reviewed and corrected by expert annotators rather than taken as ground truth.
Videos 1
Watch presentations, demos, and related content
Presentation introducing the AutoNorm Project and sharing its findings.
Like, comment, and subscribe on YouTube to support the creator!
Presentation introducing the AutoNorm Project and sharing its findings.
Gallery 1
Explore the visual story of this exhibit
Title Image