A Few‐Shot Learning Approach for a Multilingual Agro‐Information Question Answering System
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
ABSTRACT Across numerous households in Sub‐Saharan Africa, agriculture plays a crucial role. One solution that can effectively bridge the support gap for farmers in the local community is a question–answer system based on agricultural expertise and agro‐information. The more recent advancements in question answering research involve the use of large language models that are trained on an extensive amount of data. Due to this, conventional fine‐tuning approaches have demonstrated a significant decline in performance when using a significantly smaller amount of data. One proposed alternative to address this decline is to use prompt‐based fine‐tuning, which allows the model to be fine‐tuned with only a few examples, thus addressing the disparities between the objectives of pretraining and fine‐tuning. Extensive research has been done on these methods, specifically on text classification and not question answering. In this research, our objective was to study the feasibility of recent few‐shot learning approaches such as FewshotQA and Null‐prompting for domain‐specific agricultural data in four South African languages. We first explored creating a cross‐lingual domain‐specific extractive question answering dataset through an automated approach using the GPT model. Through exploratory data analysis, the GPT model was able to create a dataset, which requires minor improvements. We then evaluated the overall performance of the different approaches and investigated the effects of adapting these approaches to suit the new dataset. Results show these methods effectively capture semantic relationships and domain‐specific terminology but exhibit limitations, including potential biases in automated annotation and plateauing F1 scores. This highlights the need for hybrid approaches that combine artificial intelligence and human supervision. Beyond academic insights, this study has practical significance for industry, demonstrating how prompt‐based methods can help tailor AI models to specific use cases in low‐resource settings.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it