Hierarchical reinforcement learning with advantage function for entity 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
Unlike the traditional pipeline methods, the joint extraction approaches use a single model to distill the entities and semantic relations between entities from the unstructured texts and achieve better performances. A pioneering work, HRL-RE, uses a hierarchical reinforcement learning model to distill entities and relations that decompose the entire extraction process into a high-level relationship extraction and a low-level entity identification. HRL-RE makes the extraction of entities and relations more accurate while solving overlapped entities and relations to a certain extent. However, this method has not achieved satisfactory results in dealing with overlapped entities and relations in sentences. One reason is that learning a policy is usually inefficient, and the other one is the high variance of gradient estimators. In this paper, we propose a new method, Advantage Hierarchical Reinforcement Learning for Entity Relation Extraction (AHRL-ERE), which combines the HRL-RE model with a new advantage function to distill entities and relations from the structureless text. Specifically, based on the reference value of the policy function in the high-level subtask, we construct a new advantage function. Then, we combine this advantage function with the value function of the strategy in the low-level subtask to form a new value function. This new value function can immediately evaluate the current policy, so our AHR-ERE method can correct the direction of the policy gradient update in time, thereby making policy learning efficient. Moreover, our advantage function subtracts the reference value of the high-level policy value function from the low-level policy value function so that AHRL-ERE can decrease the variance of the gradient estimator. Thus our AHRL-ERE method is more effective for extracting overlapped entities and relations from the unstructured text. Experiments on the diffusely used datasets demonstrate that our proposed algorithm has better manifestation than the existing approaches do.
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