One Wug, Two Wug+s Transformer Inflection Models Hallucinate Affixes
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
Data augmentation strategies are increasingly important in NLP pipelines for low-resourced and endangered languages, and in neural morphological inflection, augmentation by so called data hallucination is a popular technique. This paper presents a detailed analysis of inflection models trained with and without data hallucination for the low-resourced Canadian Indigenous language Gitksan. Our analysis reveals evidence for a concatenative inductive bias in augmented models---in contrast to models trained without hallucination, they strongly prefer affixing inflection patterns over suppletive ones. We find that preference for affixation in general improves inflection performance in "wug test" like settings, where the model is asked to inflect lexemes missing from the training set. However, data hallucination dramatically reduces prediction accuracy for reduplicative forms due to a misanalysis of reduplication as affixation. While the overall impact of data hallucination for unseen lexemes remains positive, our findings call for greater qualitative analysis and more varied evaluation conditions in testing automatic inflection systems. Our results indicate that further innovations in data augmentation for computational morphology are desirable.
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 | 0.009 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.003 | 0.004 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.004 | 0.001 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.022 | 0.001 |
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