Recall Aspects of Transformers for Text Ranking
Bibliographic record
Abstract
This paper documents the University of Amsterdam’s participation in the TREC 2021 Deep Learning Track. In addition to providing labeled training data at scale, the other major contribution of the TREC DL track is to avoid the pool-bias exhibited in all the earlier adhoc search test collections created through the pooling only runs from traditional sparse retrieval systems. However, even in the TREC deep learning track, we have shallow pools, and runs with varying but high fractions of unjudged documents. This prompts a deeper analysis of pool coverage over ranks for both a representative traditional approach (i.e., BM25) and a representative neural approach (i.e., the BERT cross-encoder for the passage retrieval task). Our main conclusions are the following. First, we submitted a neural run that specifically looks beyond those documents easily found by traditional models, highlighting the potential of neural models to address recall aspects in addition to the precision aspects prioritized in the TREC Deep Learning Track up to now. Second, we observe high fractions of unjudged documents after the initial ranks for both the 2020 and 2021 data, which may hinder the evaluation of recall-oriented aspects and reusability of the judgments for runs not contributing to the pooling. Third, we observe a gradual decline of the fraction of relevant over judged documents for 2020, which is a positive sign<br/>against pooling bias, but almost no decrease for 2021. Our general conclusion is that coverage below the guaranteed pooling horizon is far from complete and that analysis of recall aspects must be done with care, but that there is great potential to study these in future editions of the track.<br/>
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 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 | 0.001 |
| Open science | 0.001 | 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 itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, 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".