Extending Retrieval in Terrier

Altering the retrieval process

It is very easy to alter the retrieval process in Terrier, as there are many hooks at which external classes can be involved. Firstly, you are free when writing your own application to render the results from Terrier in your own way. Results in Terrier come in the form of a ResultSet.

An application’s interface with Terrier is through the Manager class. The manager firstly pre-processes the query, by applying it to the configured TermPipeline. Then it calls the Matching class, which is responsible for matching documents to the query, and scoring the documents using a WeightingModel. Internally, Matching implementations use the PostingListManager to open an IterablePosting for each query term.

The terms used for matching are expressed as matching operators (matchops). It is possible to express queries as low-level matchops, e.g. using the -m option of the batchretrieve command.

The overall score of a document to the entire query can be modified by using a DocumentScoreModifier, which can be set by the matching.dsms property.

Once the ResultSet has been returned to the Manager, there are two further phases, namely PostProcessing and PostFiltering. In PostProcessing, the ResultSet can be altered in any way - for example, QueryExpansion expands the query, and then calls Matching again to generate an improved ranking of documents. PostFiltering is simpler, allowing documents to be either included or excluded - this is ideal for interactive applications where users want to restrict the domain of the documents being retrieved.

Changing Batch Retrieval

TRECQuerying is the main way in which retrieval is deployed for batch retrieval experiments. It has a multitude of ways in which it can be extended:

  • Format of input topics: Terrier supports topics in two formats (TREC tagged, or one query per line). If neither of these is suitable, then you can implement another QuerySource that knows how parse your topics files. Use the trec.topics.parser property to configure Terrier to use your new QuerySource. E.g. trec.topics.parser=my.package.DBTopicsSource.
  • Format of output results: You can implement another OutputFormat to change the format of the results in the .res files. Use the trec.querying.outputformat property to configure Terrier to use your new OutputFormat. E.g. trec.querying.outputformat=my.package.MyTRECResultsFormat.

Altering query expansion

QueryExpansion has various ways in which it can be extended:

  • To change the exact formula used to score occurrences, implement QueryExpansionModel.
  • Currently, terms are weighted from the entire feedback set as one a bag of words over the feedback set. To change this, extend ExpansionTerms.
  • To change the way feedback documents are selected, implement FeedbackSelector.

Advanced Weighting Models

It is very easy to implement your own weighting models in Terrier. Simply write a new class that extends WeightingModel. What’s more, there are many examples weighting models in org.terrier.matching.models.

Generic Divergence From Randomness (DFR) Weighting Models

The DFRWeightingModel class provides an interface for freely combining different components of the DFR framework. It breaks a DFR weighting model into three components: the basic model for randomness, the first normalisation by the after effect, and term frequency normalisation. Details of these three components can be found from a description of the DFR framework. The DFRWeightingModel class provides an alternate and more flexible way of using the DFR weighting models in Terrier. For example, to use the PL2 model, the name of the model PL2 should be given in etc/trec.models, or set using the property trec.model. Alternatively, using the DFRWeightingModel class, we can replace PL2 with DFRWeightingModel(P, L, 2), where the three components of PL2 are specified in the brackets, separated by commas. If we do not want to use one of the three components, for example the first normalisation L, we can leave the space for this component blank (i.e. DFRWeightingModel(P, , 2)). We can also discard term frequency normalisation by removing the 2 between the brackets (i.e. DFRWeightingModel(P, , )). However, a basic randomness model must always be given.

The basic randomness models, the first normalisation methods, and the term frequency normalisation methods are included in packages org.terrier.matching.models.basicmodel, org.terrier.matching.models.aftereffect and org.terrier.matching.models.normalisation, respectively. Many implementations of each are provided, allowing a vast number of DFR weighting models to be generated.

Matching strategies

Terrier implements three main alternatives for matching documents for a given query, each of which implements the Matching interface:

  • Document-At-A-Time (DAAT) (as per daat.Full) - exhaustive Matching strategy that scores all matching query terms for a document before moving onto the next documemt. Using daat.Full is advantageous for retrieving from large indices, and is the default matching strategy in Terrier.
  • Term-At-A-Time (TAAT) (as per taat.Full) - exhaustive Matching strategy that scores all postings for a single query term, before moving onto the next query term. for large indices, taat.Full consumes excessive memory with large partial result sets.
  • TRECResultsMatching - retrieves results from a TREC result file rather than the current index, based on the query id. Such a result file must be compatible with trec_eval. TRECResultsMatching can introduce a repeatable efficiency gain for batch experiments.

If you have a more complex document weighting strategy that cannot be handled as a WeightingModel or DocumentScoreModifier, you may wish to implement your own Matching strategy. In particular, BaseMatching is a useful base class. Moreover, the PostingListManager should be used for opening the IterablePosting posting stream for each query term.

Learning to Rank

Terrier support the application of learning to rank techniques within Terrier’s ranking process. This is described separately in the learning to rank documentation.

Using Terrier Indices in your own code

How many documents does term X occur in?

Index index = Index.createIndex();
Lexicon<String> lex = index.getLexicon();
LexiconEntry le = lex.getLexiconEntry("term");
if (le != null)
	System.out.println("Term term occurs in "+ le.getDocumentFrequency() + " documents");
else
	System.out.println("Term term does not occur");

What is the probability of term Y occurring in the collection?

Index index = Index.createIndex();
Lexicon<String> lex = index.getLexicon();
LexiconEntry le = lex.getLexiconEntry("X");
double p = le == null
	?  0.0d
	: (double) le.getFrequency() / index.getCollectionStatistics().getNumberOfTokens();

What terms occur in the 11th document?

Index index = Index.createIndex();
PostingIndex<?> di = index.getDirectIndex();
DocumentIndex doi = index.getDocumentIndex();
Lexicon<String> lex = index.getLexicon();
int docid = 10; //docids are 0-based
IterablePosting postings = di.getPostings(doi.getDocumentEntry(docid));
//NB: postings will be null if the document is empty
while (postings.next() != IterablePosting.EOL) {
	Map.Entry<String,LexiconEntry> lee = lex.getLexiconEntry(postings.getId());
	System.out.print(lee.getKey() + " with frequency " + postings.getFrequency());
}

What documents does term “Z” occur in, and at what position?

Index index = Index.createIndex();
PostingIndex<?> inv = index.getInvertedIndex();
MetaIndex meta = index.getMetaIndex();
Lexicon<String> lex = index.getLexicon();
LexiconEntry le = lex.getLexiconEntry( "Z" );
IterablePosting postings = inv.getPostings((BitIndexPointer) le);
while (postings.next() != IterablePosting.EOL) {
	String docno = meta.getItem("docno", postings.getId());
	System.out.println(docno + " with frequency " + postings.getFrequency());
}

If we assume that the index contains positional information, the loop above would change as follows:

while (postings.next() != IterablePosting.EOL) {
	String docno = meta.getItem("docno", postings.getId());
	int[] positions = ((BlockPosting)postings).getPositions();
	System.out.println(docno + " with frequency " + postings.getFrequency() + " and positions " + Arrays.toString(positions));
}

Moreover, if you’re not comfortable with using Java, you can dump the indices of a collection using the –print* options of the indexutil command. See the javadoc of IndexUtil for more information.

What are the PL2 weighting model scores of documents that “Y” occurs in?

Use of a WeightingModel class needs some setup, namely the EntryStatistics of the term (obtained from the Lexicon), as well as the CollectionStatistics (obtained from the index).

Index index = Index.createIndex();
PostingIndex<?> inv = index.getInvertedIndex();
MetaIndex meta = index.getMetaIndex();
Lexicon<String> lex = index.getLexicon();
LexiconEntry le = lex.getLexiconEntry( "Y" );
WeightingModel wmodel = new PL2();
wmodel.setCollectionStatistics(index.getCollectionStatistics());
wmodel.setEntryStatistics(le);
wmodel.setKeyFrequency(1);
wmodel.prepare()
IterablePosting postings = inv.getPostings((BitIndexPointer) le);
while (postings.next() != IterablePosting.EOL) {
	String docno = meta.getItem("docno", postings.getId());
	System.out.println(docno + " with score " + wmodel.score(postings));
}

Example Querying Code

Below, you can find a example sample of using the querying functionalities of Terrier.

	IndexRef indexref = IndexRef.of("/path/to/data.properties");
	Manager queryingManager = ManagerFactory.from(indexref);
	String query = "term1 term2";
	SearchRequest srq = queryingManager.newSearchRequestFromQuery(query);
	srq.addMatchingModel("org.terrier.matching.daat.Full", "PL2");
	queryingManager.runSearchRequest(srq);
	List<ScoredDoc> rs = srq.getResults();

Webpage: http://terrier.orgContact: School of Computing ScienceCopyright (C) 2004-2020 University of Glasgow. All Rights Reserved.