South African Tuberculosis Vaccine Initiative
UCT Computer Science Research Document ArchiveNot a member yet
1270 research outputs found
Sort by
Benchmarking IsiXhosa Automatic Speech Recognition and Machine Translation for Digital Health Provision
Full text linkAs digital health becomes more ubiquitous, people from different geographic regions are connected and there is thus a need for accurate language translation services. South Africa presents opportunity and need for digital health innovation, but implementing indigenous translation systems for digital health is difficult due to a lack of language resources. Understanding the accuracy of current models for use in medical translation of indigenous languages is crucial for designers looking to build quality digital health solutions. This paper presents a new dataset with audio and text of primary health consultations for automatic speech recognition and machine translation in South African English and the indigenous South African language of isiXhosa. We then evaluate the performance of well-established pretrained models on this dataset. We found that isiXhosa had limited support in speech recognition models and showed high, variable character error rates for transcription (26-70%). For translation tasks, Google Cloud Translate and ChatGPT outperformed the other evaluated models, indicating large language models can have similar performance to dedicated machine translation models for low-resource language translation
Energy Costs and Neural Complexity Evolution in Changing Environments
Full text linkThe Cognitive Buffer Hypothesis (CBH) posits that larger
brains evolved to enhance survival in changing conditions.
However, larger brains also carry higher energy demands,
imposing additional metabolic burdens. Alongside brain
size, brain organization plays a key role in cognitive ability
and, with suitable architectures, may help mitigate energy
challenges. This study evolves Artificial Neural Networks
(ANNs) used by Reinforcement Learning (RL) agents to investigate
how environmental variability and energy costs influence
the evolution of neural complexity, defined in terms
of ANN size and structure. Results indicate that under energy
constraints, increasing seasonality led to smaller ANNs.
This challenges CBH and supports the Expensive Brain Hypothesis
(EBH), as highly seasonal environments reduced net
energy intake and thereby constrained brain size. ANN structural
complexity primarily emerged as a byproduct of size,
where energy costs promoted the evolution of more efficient
networks. These results highlight the role of energy constraints
in shaping neural complexity, offering in silico support
for biological theory and energy-efficient robotic design
On the usage of semantics, syntax, and morphology for noun classification in isiZulu
Full text linkThere is limited work aimed at solving the core task of noun classification for Nguni languages. The task focuses on identifying the semantic categorisation of each noun and plays a crucial role in the ability to form semantically and morphologically valid sentences. The work by Byamugisha (2022) was the first to tackle the problem for a related, but non-Nguni, language. While there have been efforts to replicate it for a Nguni language, there has been no effort focused on comparing the technique used in the original work vs. contemporary neural methods or a number of traditional machine learning classification techniques that do not rely on human-guided knowledge to the same extent. We reproduce Byamugisha (2022)’s work with different configurations to account for differences in access to datasets and resources, compare the approach with a pre-trained transformer-based model, and traditional machine learning models that rely on less human-guided knowledge. The newly created data-driven models outperform the knowledge-infused models, with the best performing models achieving an F1 score of 0.97
IsiZulu noun classification based on replicating the ensemble approach for Runyankore
Full text linkA noun’s class is a crucial component in NLP, because it governs agreement across the sentence in Niger Congo B (NCB) languages, among others. There is a lack of computational
models for determining a noun’s class owing to ill-documentation in most NCB languages. A promising approach by Byamugisha (2022) used a data-driven approach for Runyankore
that combined syntax and semantics. The code and data are inaccessible however, and it remains to be seen whether it is suitable for other NCB languages. We solve the problem by reproducing Byamugisha’s experiment, but then for isiZulu. We conducted this as two independent experiments, so that we also could subject it to a meta-analysis. Results showed that it was reproducible only in part, mainly due to imprecision in the original description, and the current impossibility to generate the same kind of source data set generated from an existing grammar. The different choices made in attempting to reproduce the pipeline as well as differences in choice of training and test data had a large effect on the eventual accuracy of noun class disambiguation but could produce an accuracy of 83%, in the same range as Runyankore
BabyLMs for isiXhosa: Data-Efficient Language Modelling in a Low-Resource Context
Full text linkThe BabyLM challenge called on participants to develop sample-efficient language models. Submissions were pretrained on a fixed English corpus, limited to the amount of words children are exposed to in development (<100m). The challenge produced new architectures for data-efficient language modelling, outperforming models trained on trillions of words. This is promising for low-resource languages, where available corpora are limited to much less than 100m words. In this paper, we explore the potential of BabyLMs for low-resource languages, using the isiXhosa language as a case study. We pretrain two BabyLM architectures, ELC-BERT and MLSM, on an isiXhosa corpus. They outperform a vanilla pretrained model on POS tagging and NER, achieving notable gains (+3.2 F1) for the latter. In some instances, the BabyLMs even outperform XLM-R. Our findings show that data-efficient models are viable for low-resource languages, but highlight the continued importance, and lack of, high-quality pretraining data. Finally, we visually analyse how BabyLM architectures encode isiXhosa
Transformative Narratives: Fostering Ubuntu-Inspired Participatory Design Practices
Full text linkEmerging within the diverse tapestry of African wisdom and knowledge, Ubuntu, a philosophy of shared humanity and interconnectedness, offers exciting pathways for transforming participatory design (PD) practices in the Global South. Beyond a mere call for inclusion, Ubuntu inspires decolonial approaches and has the potential for sustainable innovation in design that resonates with local contexts. While communal and indigenous philosophies have increasingly guided decolonial PD efforts, particularly in the Global South, Ubuntu presents a unique lens for embracing both practical methods and a shift towards interconnectedness as a design ethos. This exploratory paper proposes concrete Ubuntu-guided PD techniques, leveraging storytelling and shared narratives, to cultivate a deep sense of interconnectedness within PD research. By amplifying these modes of engagement, we pave the way for alternative, “otherwise” knowledge production and participatory approaches, and future explorations of other indigenous and decolonial philosophies in PD research
A Human Behavior Exploration Approach Using LLMs for Cyber-Physical Systems
No full textIn the early phases of Cyber-Physical Systems (CPS) development, scoping human behavior plays a significant role, especially when interactions extend beyond expected behavior. Here, it is especially challenging to develop cases that capture the full spectrum of hu-
21 man behavior. Up to now, identifying such behavior of humans remains a task for domain experts. We explore how one can use Large Languages Models (LLMs) in the design phase of systems to provide additional information about human-CPS interaction. Our approach proposes a preliminary ontology describing a hierarchy of types of behavior and relevant CPS components as input for prompt templates. It uses them to generate parts of human behavior descriptions, as well as a canned prompt with one variable about behavior. For demonstration, we take a smart building with a Home Energy System as the use case.
31 An initial user evaluation shows that the behavior descriptions
32 generated with standard and ontology-driven prompts complement
33 each other and are useful when assisting humans. The discovered
34 uncommon behaviors can be used to complete interaction scenarios
35 that eventually result in a more robust CPS implementation
Deep Learning for Cleaning Cultural Heritage Point Clouds
Full text linkLaser scanning technology is often used in the Cultural Heritage
domain to capture the 3D structure of a site, with each scan consisting
of a set of 3D point coordinates — a point cloud. Before these
point clouds can be utilized to build a complete 3D surface model, unwanted
points must be removed. While manual point cloud cleaning is
time-consuming, previous work has shown promise in automating parts of
the process. This study builds on a previous approach which interprets
point cloud cleaning as a segmentation task accomplished via binary
point classification, applied to individual point clouds. This approach
uses a basic Random Forest (RF) classifier with hand-crafted features,
is designed to clean scans one by one via incremental per scan training,
and requires a complex graph-based post-processing step to achieve
acceptable results. By contrast, we leverage modern point-based deep
learning models to directly learn useful features, and develop a framework
that processes the fully registered set of point clouds, rather than
cleaning scans individually. Our method focuses on purely geometric attributes,
uses a few-shot fine-tuning approach and, unlike the single scan
method, does not require segmentation post-processing to improve results.
Under this scheme, users label 2.5 − 50% of an unlabelled scan,
and a model is trained to label the rest. We assess three deep learning
point-based models (Pointnet++, KPConv, Point Transformer) along
with a baseline Random Forest model, focusing on speed, accuracy, and
the reduction of total labelling effort. Our findings reveal that modern
deep learning requires minimal human labelling, with up to 85% reduction
in total labelling effort. KPConv stands out for its efficiency with
less human input, while Random Forests work best for simpler scenes.
This study highlights deep learning’s effectiveness in reducing manual
labour in point cloud cleaning in the cultural heritage domain
Multi-Objective Evolution for Chemical Product Design
Full text linkThe design of chemical products requires the optimization of desired
properties in molecular structures. Traditional techniques
are based on laboratory experimentation and are hindered by the
intractable number of alternatives and limited capabilities to identify
feasible molecules and either test or infer their properties for
optimization. Computational techniques based on deep learning
and multi-objective evolutionary optimization have spurred chemical
product design, but the definition of appropriate metrics to
compare techniques is challenging. We suggest the adoption of two
complementary assessments to account for quantitative as well
as qualitative features of different techniques, and then test our
proposed assessments by comparing two heuristics to build new
generations of molecular candidates, termed respectively, direct
correlation and extended search
Formation and immunological evaluation of Moraxella catarrhalis glycoconjugates based on synthetic oligosaccharides
No full textPublished work has shown that glycoconjugate vaccines, based on truncated detoxified lipopolysaccharides from Moraxella catarrhalis attached through their reducing end to a carrier protein, gave good protection for all three serotypes A, B, and C in mice immunisation experiments. The (from the non-reducing end) truncated LPS structures were obtained from bacterial glycosyl transferase knock-out mutants and contained the de-esterified Lipid A, two Kdo residues and five glucose moieties. This work describes the chemical synthesis of the same outer Moraxella LPS structures, spacer-equipped and further truncated from the reducing end, i.e., without the Lipid A part and containing four or five glucose moieties or four glucose moieties and one Kdo residue, and their subsequent conjugation to a carrier protein via a five‑carbon bifunctional spacer to form glycoconjugates. Immunisation experiments both in mice and rabbits of these gave a good antibody response, being 2–7 times that of pre-immune sera. However, the sera produced only recognized the immunizing glycan immunogens and failed to bind to native LPS or whole bacterial cells. Comparative molecular modelling of three alternative antigens shows that an additional (2 → 4)-linked Kdo residue, not present in the synthetic structures, has a significant impact on the shape and volume of the molecule, with implications for antigen binding and cross-reactivity