NLI-based Filtering for Data Augmentation in Topic Classification
Bibliographic record
Abstract
As manually labelling data can be labour intensive, some research proposes to automatically expand the training data in the hope of improving model's generalization ability, known as data augmentation (DA). In Natural Language Processing (NLP), some recent studies leverage pre-trained language models (PLMs) to generate new training samples and then fine-tune the underlying model to fit for downstream tasks. However, a direct use of such models tend to trigger large discrepancies and most existing post-processing methods do not work well in real-world applications. In this paper, we meet this challenge by adopting the generate-and-filter framework that consists of two phases: a generation phase which uses PLMs to develop new candidate samples without fine-tuning, and a filtering phase that determines which samples are of a better quality and could be included in the augmentation dataset. Another key contributions of our work is to formulate the filtering phase as a natural language inference (NLI) task, which makes it possible to employ sophisticated methods such as auto-encoder-based PLMs to evaluate the importance of the candidate sample. In addition, we investigate the particular choices of models used in this framework, and discuss optimization options. Experiments for the task of topic classification on three public datasets prove the effectiveness of our proposal.
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.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.012 | 0.033 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.003 |
| Bibliometrics | 0.003 | 0.003 |
| Science and technology studies | 0.002 | 0.002 |
| Scholarly communication | 0.002 | 0.004 |
| Open science | 0.003 | 0.003 |
| Research integrity | 0.002 | 0.004 |
| Insufficient payload (model declined to judge) | 0.004 | 0.003 |
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 itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".