Simplicity and learning to distinguish arguments from modifiers
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
We present a learnability analysis of the argument-modifier distinction, asking whether there is information in the distribution of English constituents that could allow learners to identify which constituents are arguments and which are modifiers. We first develop a general description of some of the ways in which arguments and modifiers differ in distribution. We then identify two models from the literature that can capture these differences, which we call the argument-only model and the argument-modifier model. We employ these models using a common learning framework based on two simplicity biases which tradeoff against one another. The first bias favors a small lexicon with highly reusable lexical items, and the second, opposing, bias favors simple derivations of individual forms – those using small numbers of lexical items. Our first empirical study shows that the argument-modifier model is able to recover the argument-modifier status of many individual constituents when evaluated against a gold standard. This provides evidence in favor of our general account of the distributional differences between arguments and modifiers. It also suggests a kind of lower bound on the amount of information that a suitably equipped learner could use to identify which phrases are arguments or modifiers. We then present a series of analyses investigating how and why the argument-modifier model is able to recover the argument-modifier status of some constituents. In particular, we show that the argumentmodifier model is able to provide a simpler description of the input corpus than the argument-only model, both in terms of lexicon size, and in terms of the complexity of individual derivations. Intuitively, the argument-modifier model is able to do this because it is able to ignore spurious modifier structure when learning the lexicon. These analyses further support our general account of the differences between arguments and modifiers, as well as our simplicity-based approach to learning.
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 | 0.001 | 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