Skip to content

Latest commit

 

History

History
95 lines (75 loc) · 6.34 KB

retrieve.md

File metadata and controls

95 lines (75 loc) · 6.34 KB
Raw

Batch query processing

The retrieve tool is used to perform batch query processing experiments. To use it:

./target/bin/retrieve

Configuration

The configuration of this script is completely managed by Java properties, as reported in the MatchingConfiguration class.

Property name Description Default
micro.namespace default namespace to append to class names not including any namespace it.cnr.isti.hpclab.matching.
micro.index.path the name of the directory in which the index data structures to process are stored .
micro.index.prefix filename prefix for the index data structures to process data
micro.manager class name of the manager to use at runtime, created by the Querying class micro.namespace + Manager
micro.model class name of the weighting model to use at runtime, create by the Manager class micro.namespace + structures.BM25
micro.matching class name of the matching algorithm to use at runtime, create by the Manager class micro.namespace + RankedOr
micro.queries comma-separated list of text file where queries are read from class QuerySource
micro.queries.id boolean value specifying if query lines contain an initial query id false
micro.queries.tokenise boolean value specifying if queries must be tokenised or treated as single terms true
micro.queries.lowercase boolean value specifying if queries must be lowercased true
micro.queries.num number of queries to process all
micro.queries.threshold file containing the thresholds for priming. If queries have id, thresholds must have ids too ""
micro.termpipelines term processors to be applied in order to single terms Stopwords,PorterStemmer
micro.ignore.low.idf boolean value specifying if terms with low IDF must be ignored true
micro.topk the number of top documents to return, if necessary 1000

The values can be changed by using Java command-line system property values:

    java -Dmicro.queries=/tmp/somefile.txt <...>

or by including a micro.properties file in the classpath.

Query Format

Queries are read by the QuerySource class, one per line, verbatim, from the file(s) specified by the micro.queries property. Empty lines and lines starting with # are ignored. By default, queries are tokenised by this class, and are passed verbatim to the query parser. Tokenisation can be turned off by the property micro.queries.tokenise. Moreover, the first token on each line can be the query id. This can be controlled by the property micro.queries.id (default: false).

Matching Algorithms

A matching algorithm must implement the MatchingAlgorithm interface. The matching algorithms currently implemented are the following.

Class Description
And Boolean AND, scores are not computed, no results are actually returned.
Or Boolean OR, scores are not computed, no results are actually returned.
RankedAnd Ranked AND processing, micro.topk documents returned with their scores.
RankedOr Ranked OR processing, micro.topk documents returned with their scores, implemented as DAAT.
MaxScore Ranked OR processing, micro.topk documents returned with their scores, implemented as Turtle & Flood's MaxScore.
Wand Ranked OR processing, micro.topk documents returned with their scores, implemented as Carmel & Broder's WAND.
BlockMaxWand Ranked OR processing, micro.topk documents returned with their scores, implemented as Suel's BlockMaxWand (postings aligned).

Alternative usage

Use the provided Java program Retrieve as follows.

java it.cnr.isti.hpclab.Retrieve [y]

The program uses the Java properties to configure its runtime. The only parameter flag (y), if present, will print to stderr all configured properties with their values, waiting for confirmation to proceed. Otherwise, the query processing will begin, outputting results to stdout in a simple JSON format. For example:

java -Xmx32G -server \
     -cp terrier-micro-1.4.0.jar \
     -Dmicro.index.path=/data1/khast/index-java \
     -Dmicro.index.prefix=cw09b.sux \
     -Dmicro.queries=./query_log/msn.1k.txt \
     -Dmicro.topk=30 \
     -Dmicro.matching=it.cnr.isti.hpclab.matching.RankedOr \
     -Dstopwords.filename=/home/khast/stuff/stopword-list.txt \
     it.cnr.isti.hpclab.Retrieve

Output

During query processing, a JSON file is produced, one line per query. Every line contains information about the terms and the processing time of the query. It can be easily parsed with the jq tool.

An additional output can be generate, if the actual docids returned for every query are necessary (typically, for debugging purposes or for effectiveness measures). This output is controlled by the following properties.

Property name Description Default
micro.results.output.type the type of output for results generation. Possible values are null, docid, score, docno and trec null
micro.results.filename the absolute filename of the gzipped file where the output results will be stored results.gz

Depending on the output type, the content of the output type will be different, according to the following descriptions.

  • null: no output is actually generated. This is the default behavior, since writing the results on file might negatively impact the efficiency of query processing.
  • docid: each line of the output file has the format <qid>TAB<docid>.
  • score: each line of the output file has the format <qid>TAB<docid>TAB<score>. The score value is ceiled to 5 decimal digits.
  • docno: each line of the output file has the format <qid>TAB<docno>TAB<score>. The score value is ceiled to 5 decimal digits. The docno is retrieved from the metaindex.
  • trec: each line of the output file has the format <qid> Q0 <docno> <pos> <score> ISTI (note the spaces instead of tabs and the 'immutable' strings Q0 and ISTI). The score value is ceiled to 5 decimal digits. The docno is retrieved from the metaindex. The pos value is the position in the returned results array, starting from 0 (i.e., the highest scoring document is in position 0, etc.).