Focused hierarchical RNNs for conditional sequence processing
Bibliographic record
Abstract
Recurrent Neural Networks (RNNs) with atten-tion mechanisms have obtained state-of-the-artresults for many sequence processing tasks. Mostof these models use a simple form of encoderwith attention that looks over the entire sequenceand assigns a weight to each token indepen-dently.We present a mechanism for focus-ing RNN encoders for sequence modelling taskswhich allows them to attend to key parts of theinput as needed. We formulate this using a multi-layer conditional sequence encoder that reads inone token at a time and makes a discrete deci-sion on whether the token is relevant to the con-text or question being asked. The discrete gatingmechanism takes in the context embedding andthe current hidden state as inputs and controls in-formation flow into the layer above. We train itusing policy gradient methods. We evaluate thismethod on several types of tasks with differentattributes. First, we evaluate the method on syn-thetic tasks which allow us to evaluate the modelfor its generalization ability and probe the behav-ior of the gates in more controlled settings. Wethen evaluate this approach on large scale Ques-tion Answering tasks including the challengingMS MARCO and SearchQA tasks. Our mod-els shows consistent improvements for both tasksover prior work and our baselines. It has alsoshown to generalize significantly better on syn-thetic tasks as compared to the baselines.
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 | 0.001 | 0.003 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.008 | 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".