[Previous: Configure Retrieval] [Contents] [Next: Query Language]

Learning to Rank with Terrier

Since version 4.0, Terrier supports the deployment of many retrieval features, and integration with learning to rank techniques. This page explains how to configure Terrier to enable learning and application of a learned model, while a worked example using the TREC .GOV corpus is provided below.

Introduction

Learning to rank is the ability to (a) use multiple features in a uniform way during ranking, and (b) to learn an appropriate method to combine those features. For learning, Terrier has the ability to calculate multiple features, be they query-dependent (c.f. multiple weighting models, such as BM25, PL2 etc), or be they query-independent (c.f. Document Length).

If you wish to use Terrier with a learning-to-rank technique, you must generate a LETOR-formatted file -- which can be done using Terrier's Normalised2LETOROutputFormat class -- to provide as to a learning-to-rank technique such as Jforests. A LETOR formatted file looks as follows:

#1: featureName
#2: featureName
0 qid:1 1:2.9 2:9.4 # docid=clueweb09-00-01492
At the top is an optional comment header giving the names of the features. Then, each line has a 'label', i.e. derived from relevance assessments such as TREC qrel files, the query Id, and featureId-featureValue pairs. Finally, each line can have the docno.

Fat Component

What is "Fat" about? Fat is a method of allowing many features to be computed within one run of Terrier. In particular, computing every feature for every posting of every query term is very expensive, and in practice, unnecessary. (Indeed learning to rank becomes slower and less accurate as the number of documents per query increases). Instead, Liu [1] suggest ranking a sample of documents using a simple weighting model (e.g. BM25) is sufficient, before computing the features on the documents in the sample.

However, once the sample has been identified, the posting lists have been iterated, and it is no longer possible to compute the other weighting model features. The Fat component [2] addresses this problem, by storing copies of the postings for every document that makes the top k retrieved documents. These can then be later used to calculate other features.

NB: If you use the Fat component in your research, you should cite [2].

Moreover, in addition to the deployed features, the number of documents to retrieve in the sample is a key parameter - for a study of this parameter in Web search, see [3].

In the following, we firstly define the classes related to Fat within the learning component of Terrier, before showing how they can be combined in pipelines for various applications.

Fat Classes

Example Using .GOV corpus - Named Page Retrieval

In the following, we give an example of effective retrieval using learning to rank, using the topics and qrels from the named page tasks of the TREC 2002-2004 Web tracks. Other possible invocations of the Fat framework are discussed below. Firstly, we setup the Terrier environment. We use simple variables in the form of a Unix Bash shell script, but this could be easily ported to a Windows environment.
#available from http://ir.dcs.gla.ac.uk/test_collections/access_to_data.html 
CORPUS=/extra/Collections/TREC/GOV/

#available from http://trec.nist.gov/data/webmain.html
TR_TOPICS=/extra/TopicsQrels/TREC/GOV/namedpage/TREC2002/webnamed_page_topics.1-150.txt
VA_TOPICS=/extra/TopicsQrels/TREC/GOV/namedpage/TREC2003/topics.NP151-NP450.np
TE_TOPICS=/extra/TopicsQrels/TREC/GOV/namedpage/TREC2004/topics.WT04-1-WT04-225.np

TR_QRELS=/extra/TopicsQrels/TREC/GOV/namedpage/TREC2002/qrels.named-page.txt
VA_QRELS=/extra/TopicsQrels/TREC/GOV/namedpage/TREC2003/qrels.NP151-NP450.np
TE_QRELS=/extra/TopicsQrels/TREC/GOV/namedpage/TREC2004/qrels.WT04-1-WT04-225.np

Firstly, we setup Terrier, and, if necessary, index the corpus using blocks, and saving field information for the TITLE and body fields:

bin/trec_setup.sh $CORPUS

echo block.indexing=true >> etc/terrier.properties
echo FieldTags.process=TITLE,ELSE >> etc/terrier.properties

bin/trec_terrier.sh -i -j

Setting TERRIER_HOME to /home/terrier-4.0
INFO - TRECCollection read collection specification (4613 files)
INFO - Processing /extra/Collections/TREC/DOTGOV/G00/02.gz
INFO - Indexer using 2 fields
Starting building the inverted file (with blocks)...
...
INFO - Collection #0 took 3835 seconds to build the runs for 1247753 documents
INFO - Merging 2 runs...
INFO - Collection #0 took 151 seconds to merge
INFO - Collection #0 total time 3986
INFO - Optimising structure lexicon
INFO - lexicon has 2 fields
INFO - Optimising lexicon with 2759934 entries
INFO - All ids for structure lexicon are aligned, skipping .fsomapid file
Finished building the inverted index...
Time elapsed for inverted file: 3989.925
Time elapsed: 3990.343 seconds.

Next, we wish to configure retrieval. We will use the Fat framework to retrieve 1000 documents using the DPH weighting model, and then calculate several additional query dependent and query independent features:

#configure the features list file
cat <<EOF > etc/features.list
WMODEL:SingleFieldModel(BM25,0)
WMODEL:SingleFieldModel(BM25,1)
QI:SingleFieldModel(Dl,0)
QI:SingleFieldModel(Dl,1)
DSM:org.terrier.matching.dsms.DFRDependenceScoreModifier
DSM:org.terrier.matching.dsms.MRFDependenceScoreModifier
EOF

Next, we want to retrieve results for the training topics. In this, we're going to be calculating results with multiple features, as listed in the etc/features.list file, so we use a series of Matching classes: FatFull to make a FatResultSet (i.e. a ResultSet with extra posting information), and FatFeaturedScoringMatching to add the additional features, and return a FeaturedResultSet. We then add the document label from the qrels using LabelDecorator, and write the results in a LETOR-compatible results file using Normalised2LETOROutputFormat:

bin/trec_terrier.sh -r -Dtrec.model=DPH -Dtrec.topics=$TR_TOPICS -Dtrec.matching=FatFeaturedScoringMatching,org.terrier.matching.daat.FatFull -Dfat.featured.scoring.matching.features=FILE -Dfat.featured.scoring.matching.features.file=$PWD/etc/features.list  -Dtrec.querying.outputformat=Normalised2LETOROutputFormat -Dquerying.postprocesses.order=QueryExpansion,org.terrier.learning.LabelDecorator -Dquerying.postprocesses.controls=labels:org.terrier.learning.LabelDecorator,qe:QueryExpansion -Dquerying.default.controls=labels:on -Dlearning.labels.file=$TR_QRELS -Dtrec.results.file=tr.letor -Dproximity.dependency.type=SD


Setting TERRIER_HOME to /home/terrier-4.0
INFO - Structure meta reading lookup file into memory
INFO - Structure meta loading data file into memory
INFO - time to intialise index : 1.113
INFO - NP1 : america s century farms
INFO - Processing query: NP1: 'america s century farms'
INFO - Query NP1 with 3 terms has 3 posting lists
term america ks=1.0 es=term339011 Nt=94640 TF=229637 @{0 500765856 0} TFf=3801,225836
term centuri ks=1.0 es=term598109 Nt=39494 TF=81231 @{0 662970754 0} TFf=699,80532
term farm ks=1.0 es=term983771 Nt=40908 TF=178987 @{0 1009035982 2} TFf=2488,176499
Term: america qtw=1.0 es=term339011 Nt=94640 TF=229637 @{0 500765856 0} TFf=3801,225836
Term: centuri qtw=1.0 es=term598109 Nt=39494 TF=81231 @{0 662970754 0} TFf=699,80532
Term: farm qtw=1.0 es=term983771 Nt=40908 TF=178987 @{0 1009035982 2} TFf=2488,176499
INFO - Rescoring found 222 docs with +ve score using SingleFieldModel(BM25,0)
Term: america qtw=1.0 es=term339011 Nt=94640 TF=229637 @{0 500765856 0} TFf=3801,225836
Term: centuri qtw=1.0 es=term598109 Nt=39494 TF=81231 @{0 662970754 0} TFf=699,80532
Term: farm qtw=1.0 es=term983771 Nt=40908 TF=178987 @{0 1009035982 2} TFf=2488,176499
INFO - Rescoring found 1000 docs with +ve score using SingleFieldModel(BM25,1)
ngramC=1.0
read: term0 Nt=94640 TF=229637 @{0 500765856 0} TFf=3801,225836 => 0
read: term1 Nt=39494 TF=81231 @{0 662970754 0} TFf=699,80532 => 1
read: term2 Nt=40908 TF=178987 @{0 1009035982 2} TFf=2488,176499 => 2
INFO - Query NP1 with 3 terms has 3 posting lists
phrase term: america
phrase term: centuri
phrase term: farm
DFRDependenceScoreModifier altered scores for 1000 documents
read: term0 Nt=94640 TF=229637 @{0 500765856 0} TFf=3801,225836 => 0
read: term1 Nt=39494 TF=81231 @{0 662970754 0} TFf=699,80532 => 1
read: term2 Nt=40908 TF=178987 @{0 1009035982 2} TFf=2488,176499 => 2
INFO - Query NP1 with 3 terms has 3 posting lists
phrase term: america
phrase term: centuri
phrase term: farm
MRFDependenceScoreModifier altered scores for 1000 documents
INFO - Applying LabelDecorator
INFO - Writing results to /home/terrier-4.0/var/results/tr.letor
...
INFO - Settings of Terrier written to /home/terrier-4.0/var/results/tr.letor.settings
INFO - Finished topics, executed 150 queries in 361.379 seconds, results written to /home/terrier-4.0/var/results/tr.letor
Lets a have a look at what was output into tr.letor:
# 1:score
# 2:WMODEL:SingleFieldModel(BM25,0)
# 3:WMODEL:SingleFieldModel(BM25,1)
# 4:QI:SingleFieldModel(Dl,0)
# 5:QI:SingleFieldModel(Dl,1)
# 6:DSM:org.terrier.matching.dsms.DFRDependenceScoreModifier
# 7:DSM:org.terrier.matching.dsms.MRFDependenceScoreModifier
1 qid:NP1 1:1.0 2:0.8508957593838526 3:1.0 4:0.38888888888888884 5:0.03218193520703712 6:0.9170723523167136 7:0.9586800382580136 #docid = 1241 docno = G00-65-2264297
1 qid:NP1 1:0.9020171593497216 2:0.8508957593838526 3:0.9169226938397537 4:0.38888888888888884 5:0.022634627762282773 6:1.0 7:1.0 #docid = 11394 docno = G00-04-3805407
0 qid:NP1 1:0.6440567566141826 2:0.0 3:0.85986678612829 4:0.0 5:0.22902810555674746 6:0.21370705023195802 7:0.3992684087689166 #docid = 921940 docno = G34-15-0261249

The header reports the name of the features. "score" means the model used to generate the sample, in our case DPH. After the header, for each retrieved document for each query, there is a single line in the output. The label obtained from the qrels file is the first entry on each row.

We repeat the retrieval step for the validation queries, this time from the 2003 TREC task:

bin/trec_terrier.sh -r -Dtrec.model=DPH -Dtrec.topics=$VA_TOPICS -Dtrec.matching=FatFeaturedScoringMatching,org.terrier.matching.daat.FatFull -Dfat.featured.scoring.matching.features=FILE -Dfat.featured.scoring.matching.features.file=$PWD/etc/features.list  -Dtrec.querying.outputformat=Normalised2LETOROutputFormat -Dquerying.postprocesses.order=QueryExpansion,org.terrier.learning.LabelDecorator -Dquerying.postprocesses.controls=labels:org.terrier.learning.LabelDecorator,qe:QueryExpansion -Dquerying.default.controls=labels:on -Dlearning.labels.file=$VA_QRELS -Dtrec.results.file=va.letor -Dproximity.dependency.type=SD

To obtain a learned model, we use the Jforests learning to rank technique, which is included with Terrier 4.0. In particular, we use Jforests data preparation command to prepare the LETOR formatted results files, then learn a LambdaMART learned model. The Jforests configuration comes entirely from the Jforests website:

bin/anyclass.sh edu.uci.jforests.applications.Runner  --config-file etc/jforests.properties --cmd=generate-bin --ranking --folder var/results/ --file tr.letor  --file va.letor

#configure the jforests file
cat  <<EOF > etc/jforests.properties
trees.num-leaves=7
trees.min-instance-percentage-per-leaf=0.25
boosting.learning-rate=0.05
boosting.sub-sampling=0.3
trees.feature-sampling=0.3
boosting.num-trees=2000
learning.algorithm=LambdaMART-RegressionTree
learning.evaluation-metric=NDCG
params.print-intermediate-valid-measurements=true
EOF

bin/anyclass.sh edu.uci.jforests.applications.Runner  --config-file etc/jforests.properties --cmd=train --ranking --folder var/results/ --train-file var/results/tr.bin --validation-file var/results/va.bin --output-model ensemble.txt
Once the learned model (from Jforests, this is an XML file which takes the form of a gradient boosted regression tree) is obtained in ensemble.txt, we can use this to apply the learned model. The configuration for Terrier is similar to retrieval for the training topics, but we additionally use JforestsModelMatching for application of the learned model, and output the final results using the default, trec_eval compatible TRECDocnoOutputFormat:
bin/trec_terrier.sh -r -Dtrec.model=DPH -Dtrec.topics=$TE_TOPICS -Dtrec.matching=JforestsModelMatching,FatFeaturedScoringMatching,org.terrier.matching.daat.FatFull -Dfat.featured.scoring.matching.features=FILE -Dfat.featured.scoring.matching.features.file=$PWD/etc/features.list -Dtrec.results.file=te.res -Dfat.matching.learned.jforest.model=$PWD/ensemble.txt -Dfat.matching.learned.jforest.statistics=$PWD/var/results/jforests-feature-stats.txt -Dproximity.dependency.type=SD
Finally, for comparison, we additionally make a simple DPH run:
bin/trec_terrier.sh -r -Dtrec.model=DPH -Dtrec.topics=$TE_TOPICS
On evaluating the two runs using trec_eval for Mean Reciprocal Rank, we find a marked increase in effectiveness, despite the deployment of no Web-specific features (such as anchor text, URL or link analysis features).
trec_eval -m recip_rank $TE_QRELS var/results/DPH_0.res
recip_rank            	all	0.4447
trec_eval  -m recip_rank $TE_QRELS var/results/te.res 
recip_rank            	all	0.5201

NB: You can equally use Terrier's evaluation tool to conduct the evaluation.

Other possible usages

In the following, we give typical configurations for using the learning/fat components of Terrier.

From inverted index -> LETOR file with many features

bin/trec_terrier.sh -r -Dtrec.matching=FatFeaturedScoringMatching,org.terrier.matching.daat.FatFull -Dfat.featured.scoring.matching.features=FILE -Dfat.featured.scoring.matching.features.file=/path/to/list.features -Dtrec.querying.outputformat=org.terrier.learning.Normalised2LETOROutputFormat

From inverted index -> Fat result file -> LETOR file with many features

You can save intermediate FatResultSets, to that can go back and compute different sets of features without retrieval from the inverted index.
bin/trec_terrier.sh -r -Dtrec.matching=org.terrier.matching.daat.FatFull -Dtrec.querying.outputformat=org.terrier.applications.WritableOutputFormat

bin/trec_terrier.sh -r -Dtrec.matching=FatFeaturedScoringMatching,FatResultsMatching -Dfat.results.matching.file=bla.fat.res.gz  -Dfat.featured.scoring.matching.features=FILE -Dfat.featured.scoring.matching.features.file=/path/to/list.features -Dtrec.querying.outputformat=org.terrier.learning.Normalised2LETOROutputFormat

From inverted index -> Final Ranking having applied learned model to documents

bin/trec_terrier.sh -r -Dtrec.matching=JforestsModelMatching,FatFeaturedScoringMatching,org.terrier.matching.daat.FatFull -Dfat.featured.scoring.matching.features=FILE -Dfat.featured.scoring.matching.features=$PWD/list.features -Dfat.matching.learned.jforest.model=/path/to/jforest.model

References

  1. Tie-Yan Lui. Learning to Rank for Information Retrieval. Foundations & Trends in Information Retrieval. 3(3):225-331, 2009.
  2. Craig Macdonald, Rodrygo L.T. Santos, Iadh Ounis and Ben He. About Learning Models with Multiple Query Dependent Features. Transactions on Information Systems. 31(3):1-39. 2013.
  3. Craig Macdonald, Rodrygo Santos and Iadh Ounis. The Whens and Hows of Learning to Rank. Information Retrieval 16(5):584-628. 2012.

[Previous: Configure Retrieval] [Contents] [Next: Query Language]