DReD–A Descriptive Relation Dataset for Expanding Relation Extraction
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
Relation extraction is a fundamental topic in document information extraction. Traditionally, datasets for relation extraction have been annotated with named entities and classified with a subset of relation categories. Models then predict either the entities and relations (end-to-end) or assume the entities are given and only classify the relations. However, current approaches are limited by datasets with a narrow definition of entities and relations. We seek to remedy this by introducing our Descriptive Relation Dataset (DReD), which contains 3286 annotations for descriptions of relations between more general noun phrases inspired by linguistic theory. We benchmark our dataset using several seq2seq models and find that T5 achieves the best results with a ROUGE-1 score of 75.5. We verify the usefulness of DreD by collecting feedback on 100 predictions and comparing human judgment to automated scoring methods. Finally, we verify that relations can be described accurately by transforming the CoNLL04 and Re-TACRED datasets and mapping sentence templates to relation categories. T5 achieves competitive accuracy on CoNLL-04 and Re-TACRED with an F1 score of 78.6 and 90.4, respectively. With this article, we prove that relations can be described, therefore overcoming the limitations set by previous datasets and approaches. We publicly provide our dataset and training code at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/logan-markewich/DReD</uri> . Relation extraction is a powerful task, providing a method to extract labeled connections between words in a document. Existing datasets focus on relations between important named entities, with relations sourced from a list of predefined categories. These categories create limitations for trained models, missing important context that a category name cannot capture alone. Our new Descriptive Relation Dataset, DReD, overcomes these limitations by providing a dataset that allows models to learn how to describe relations in a sentence. DReD contains 3286 annotations of descriptions of relations between general noun phrases, removing the previously stated limitations and providing a way to uncover previously unseen relation types while providing meaningful context. Furthermore, any sequence-to-sequence model can be easily trained on DReD, allowing for flexible and future-proof applications.
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 | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| 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