Machine Learning for Information Retrieval: TREC 2009 Web, Relevance Feedback and Legal Tracks.
Bibliographic record
Abstract
For the TREC 2009, we exhaustively classified every document in each corpus, using machine learning methods that had previously been shown to work well for email spam [9, 3]. We treated each document as a sequence of bytes, with no tokenization or parsing of tags or meta-information. This approach was used exclusively for the adhoc web, diversity and relevance feedback tasks, as well as to the batch legal task: the ClueWeb09 and Tobacco collections were processed end-to-end and never indexed. We did the interactive legal task in two phases: first, we used interactive search and judging to find a large and diverse set of training examples; then we used active learning process, similar to what we used for the other tasks, to find find more relevant documents. Finally, we fitted a censored (i.e. truncated) mixed normal distribution to estimate recall and the cutoff to optimize F1, the principal effectiveness measure. 2 Processing ClueWeb09 for Web and Relevance Feedback We used all the English documents in the full (category A) ClueWeb09 collection. The four distribution drives were mounted on a standard PC with Intel E7400 2.80GHz dual-core processor, 4GB RAM. Decompressing the 12TB of data using gzip requires about 12 hours using both cores; the learning method (for 50 topics in parallel) adds about 6 hours to this time. That is, the score for every document in the collection with respect to every topic is computed in about 18 hours.
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 | 0.043 | 0.055 |
| Meta-epidemiology (narrow) | 0.007 | 0.002 |
| Meta-epidemiology (broad) | 0.006 | 0.003 |
| Bibliometrics | 0.012 | 0.010 |
| Science and technology studies | 0.005 | 0.003 |
| Scholarly communication | 0.008 | 0.009 |
| Open science | 0.011 | 0.004 |
| Research integrity | 0.007 | 0.008 |
| Insufficient payload (model declined to judge) | 0.035 | 0.041 |
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".