TREC WT2G Collection

Below, we give an example of using Terrier, in order to index WT2G, a standard TREC test collection. We assume that the operating system is Linux, and that the collection, along with the topics and the relevance assessments, is stored in the directory /local/collections/WT2G/.

#goto the terrier folder
cd terrier

#get terrier setup for using a trec collection
bin/trec_setup.sh /local/collections/WT2G/

#rebuild the collection.spec file correctly
find /local/collections/WT2G/ -type f | sort |grep -v info > etc/collection.spec

#use In_expB2 DFR model for querying
echo uk.ac.gla.terrier.matching.models.In_expB2 > etc/trec.models

#use this file for the topics
echo /local/collections2/WT2G/info/topics.401-450.gz >> etc/trec.topics.list

#use this file for query relevance assessments
echo /local/collections2/WT2G/info/qrels.trec8.small_web.gz >> etc/trec.qrels

#index the collection
bin/trec_terrier.sh -i

#add the language modelling indices for PonteCroft
bin/trec_terrier.sh -i -l

#run the topics, with suggested c value 10.99 
bin/trec_terrier.sh -r -c 10.99
#run topics again with query expansion enabled
bin/trec_terrier.sh -r -q -c 10.99
#run topics again, using PonteCroft language modelling instead of statistical models
bin/trec_terrier.sh -r -l

#evaluate the results in var/results/
bin/trec_terrier.sh -e

#display the Mean Average Precision
tail -1 var/results/*.eval
#MAP should be 
#In_expB2 Average Precision: 0.3160 

TREC Blogs06 Collection

This guide will provide a step-by-step example on how to use Terrier for indexing, retrieval and evaluation. We use TREC Blogs06 test collection, along with the corresponding topics and the qrels from TREC 2006 Blog track. We assume that these are stored in the directory /local/collections/Blog06/

Indexing

In the Terrier folder, use trec_setup.sh to generate a collection.spec for indexing the collection:

[user@machine terrier]$ ./bin/trec_setup.sh /local/collections/Blog06/
[user@machine terrier]$ find /local/terrier/Collections/TREC/Blogs06Collection/ -type f  
	| grep 'permalinks-' |sort > etc/collection.spec

This will result in the creation of a collection.spec file, in the etc directory, containing a list of the files in the /local/collections/Blog06/ directory. At this stage, you should check the etc/collection.spec, to ensure that it only contains files that should be indexed, and that they are sorted (ie 20051206/permalinks-000.gz is the first file).

The TREC Blogs06 collection differs from other TREC collections in that not all tags should be indexed. For this reason, you should configure the parse in TRECCollection not to process these tags. Set the following properties in your etc/terrier.properties file:

TrecDocTags.doctag=DOC
TrecDocTags.idtag=DOCNO
TrecDocTags.skip=DOCHDR,DATE_XML,FEEDNO,BLOGHPNO,BLOGHPURL,PERMALINK

Finally, the length of the DOCNOs in the TREC Blogs06 collection are 30 characters, longer than the default 20 characters in Terrier. To deal with this, set the property docno.byte.length to 30 in your terrier.properties:

[user@machine terrier]$  echo docno.byte.length=30>>etc/terrier.properties

Now you are ready to start indexing the collection.

[user@machine terrier]$ ./bin/trec_terrier.sh -i
Setting TERRIER_HOME to /local/terrier
INFO - TRECCollection read collection specification
INFO - Processing /local/collections/Blogs06/20051206/permalinks-000.gz
INFO - creating the data structures data_1
INFO - Processing /local/collections/Blogs06/20051206/permalinks-001.gz
INFO - Processing /local/collections/Blogs06/20051206/permalinks-002.gz
DEBUG - flushing lexicon
<snip>

If we did not plan to use Query Expansion initially, then the faster single-pass indexing could be enabled, using the -j option of TrecTerrier. If we decide to use query expansion later, we can use the Inverted2DirectIndexBuilder to create the direct index (BlockInverted2DirectIndexBuilder for blocks).

[user@machine terrier]$ ./bin/trec_terrier.sh -i -j
Setting TERRIER_HOME to /local/terrier
INFO - TRECCollection read collection specification
INFO - Processing /local/collections/Blogs06/20051206/permalinks-000.gz
Starting building the inverted file...
INFO - creating the data structures data_1
INFO - Creating IF (no direct file)..
INFO - Processing /local/collections/Blogs06/20051206/permalinks-001.gz
INFO - Processing /local/collections/Blogs06/20051206/permalinks-002.gz
<snip>
[user@machine terrier]$ ./bin/anyclass.sh uk.ac.gla.terrier.structures.indexing.singlepass.Inverted2DirectIndexBuilder
INFO - Generating a direct index from an inverted index
INFO - Iteration - 1 of 20 iterations
INFO - Generating postings for documents with ids 0 to 120435
INFO - Writing the postings to disk
<snip>
INFO - Finishing up: rewriting document index
INFO - Finished generating a direct index from an inverted index

Indexing will take a reasonable amount of time on a modern machine. Additionally, expect to double indexing time if block indexing is enabled. Using single-pass indexing is significantly faster, even if the direct file has to be built later.

Retrieval

Once the index is built, we can do retrieval using the index, following the steps described below.

First, tell Terrier the location of the topics and relevance assessments (qrels).

[user@machine terrier]$ echo /local/collections/Blog06/06.topics.851-900 >> etc/trec.topics.list
[user@machine terrier]$ echo /local/collections/Blog06/qrels.blog06 >> etc/trec.qrels

Next, we should specify the retrieval weighting model that we want to use. In this case we will use the DFR model called PL2 for ranking documents.

echo uk.ac.gla.terrier.matching.models.PL2 > etc/trec.models

Now we are ready to start retrieval. We use the -c to set the parameter of the weighting model to the value 1. Terrier will do retrieval by taking each query (called a topic) from the specified topics file, query the index using it, and save the results to a file in the var/results folder, named similar to PL2c1.0_0.res. The file PL2c1.0_0.res.settings contains a dump of the properties and other settings used to generated the run.

[user@machine terrier]$ ./bin/trec_terrier.sh -r -c 1
Setting TERRIER_HOME to /local/terrier
INFO - 900 : mcdonalds
INFO - Processing query: 900
<snip>
INFO - Finished topics, executed 50 queries in 27 seconds, results written to 
	terrier/var/results/PL2c1.0_0.res
Time elapsed: 40.57 seconds.

Evaluation

We can now evaluate the retrieval performance of the generated run using the qrels specified earlier:

[user@machine terrier]$ ./bin/trec_terrier.sh -e
Setting TERRIER_HOME to /local/terrier
INFO - Evaluating result file: /local/terrier/var/results/PL2c1.0_0.res
Average Precision: 0.2703
Time elapsed: 3.177 seconds.

Note that more evaluation measures are stored in the file var/results/PL2c1.0_0.eval.

Common TREC Settings

This page provides examples of settings for indexing and retrieval on TREC collections. For example, to index the disk1&2 collection, the etc/terrier.properties should look like as follows:

#directory names
terrier.home=/home/me/terrier/

#default controls for query expansion
querying.postprocesses.order=QueryExpansion
querying.postprocesses.controls=qe:QueryExpansion

#default and allowed controls
querying.default.controls=c:1.0,start:0,end:999
querying.allowed.controls=c,scope,qe,qemodel,start,end

matching.retrieved_set_size=1000

#document tags specification
#for processing the contents of
#the documents, ignoring DOCHDR
TrecDocTags.doctag=DOC
TrecDocTags.idtag=DOCNO
TrecDocTags.skip=DOCHDR
#the tags to be indexed
TrecDocTags.process=TEXT,TITLE,HEAD,HL
#do not store position information in the index. Set it to true otherwise.
block.indexing=false

#query tags specification
TrecQueryTags.doctag=TOP
TrecQueryTags.idtag=NUM
TrecQueryTags.process=TOP,NUM,TITLE
TrecQueryTags.skip=DOM,HEAD,SMRY,CON,FAC,DEF,DESC,NARR

#stop-words file. default folder is ./share
stopwords.filename=stopword-list.txt

#the processing stages a term goes through
#the following setting applies standard stopword removal and Porter's stemming algorithm.
termpipelines=Stopwords,PorterStemmer

The following table lists the indexed tags (corresponding to the property TrecDocTags.process) and the running time for a singlepass inverted index creation on 6 TREC collections. No indexed tags are specified for the WT2G, WT10G, DOTGOV and DOTGOV2 collections, which means the system indexes everything in these collections. The indexing was done on a CentOS 5 Linux machine with Intel(R) Core(TM)2 2.4GHz CPU and 2G RAM (a maximum of 1G RAM is allocated to the Java virtual machine).

The following table compares the indexing time using the classical two-phase indexing (since the very first version of Terrier) and single-pass indexing (since v2.0) with and without storing the terms positions (blocks). The table shows that the single-pass indexing is markedly faster than the two-phase indexing, particular when block indexing is enabled.

The following table lists the retrieval performance achieved using three weighting models, namely the Okapi BM25, DFR PL2 and the new parameter-free DFRee model on a variety of standard TREC test collections. We provide the best values for the b and c parameters of BM25 and PL2 respectively, by optimising MAP using a simulated annealing process. In contrast, DFRee performs robustly across all collections while it does not require any parameter tuning or training.

Many of the above TREC collections can be obtained directly from either TREC (NIST), or from the University of Glasgow

Copyright © 2015 University of Glasgow | All Rights Reserved