Impact of Tokenization, Pretraining Task, and Transformer Depth on Text Ranking
Bibliographic record
Abstract
This paper documents the University of Amsterdam’s participation in the TREC 2020 Deep Learning Track. Rather than motivated by engineering the best scoring system, our work is motivated by our interest in analysis, informing our understanding of the opportunities and challenges of transformers for text ranking. Specifically, we focus on the passage retrieval task where we try to answer three of sets of questions. First, transformers use different tokenization than traditional IR approaches such as stemming and lemmatizing, leading to different document representations. What is the effect of modern preprocessing techniques on traditional retrieval algorithms? Our main observation is that the limited vocabulary of the BERT tokenizer is affecting many long-tail tokens, which leads to large gains in efficiency at the cost of a small decrease in effectiveness. Second, the effectiveness of transformers is a result of the self-supervised pre-training task promoting general language understanding, ignorant of the specific demands of ranking tasks. Can we make further correlate queries and relevant passages in the pre-training task? Our main observation is that there is a whole continuum between the original self-supervised training task of BERT and the final interaction ranker, and isolating ranking-aware pre-training tasks may leads to gains in efficiency (as these pretrained models can be reused for many tasks) as well as to gains in effectiveness (in particular when limited data on the target task is available). Third, transformers combine large sequence length with many layers, with unclear what this deep semantics adds in the context of ranking. How complex do the models need to be in order to perform well on this task? Our main observation is that the deep layers of BERT lead to some, but relatively modest, gains in performance, but that the exact role of the presumed superior language understanding for search is far from clear.
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 | 0.000 |
| 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 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".