Search Engines:
11-442 / 11-642 / 11-742

 

 

HW2 requires you to understand and make several extensions to the QryEval search engine architecture. We recommend the following development path.

Start with BM25, which is the simpler of the two retrieval models. There are five main steps.

• Create the RetrievalModelBM25 class. The ranker.BM25:xxx parameters need to find their way from the .param file into this class. The default query operator is #SUM.
• Modify the #SCORE operator to calculate scores for RetrievalModelBM25.
  • Remember to think about which parts of the BM25 score calculation are document-specific and which are document-independent. Your software will be faster if the document-independent portions are calculated just once (e.g., during query initialization) and cached somewhere in the query tree.
• Create a QrySopSum class. Copy the QrySopOr class and adjust how scores are calculated.
• Modify the QryParser class to recognize #SUM operators in queries.

Indri is more complex. It is easiest to tackle Indri in two stages. In the first stage, follow an implementation process that is similar to what was used for BM25.

• Create the RetrievalModelIndri class. The ranker.Indri:xxx parameters need to find their way from the .param file into this class. The default query operator is #AND.
• Modify the #SCORE operator to calculate scores for RetrievalModelIndri.
  • Remember to think about which parts of the Indri score calculation are document-specific and which are document-independent. Your software will be faster if the document-independent portions are calculated just once (e.g., during query initialization) and cached somewhere in the query tree.
• Add a getDefaultScore function to QrySopScore. See the HW2 Implementation slides.
• You already have a QrySopAnd class. It needs three adjustments.
  • Think about whether the logic of the matching criteria (docIteratorHasMatch) needs to be adjusted when using RetrievalModelIndri.
  • Implement the #AND operator score calculation for RetrievalModelIndri.
  • Add a getDefaultScore function. See the HW2 Implementation slides.

Once the Indri #AND operator is working properly, repeat that same process for the #WAND and #WSUM operators. There are two steps.

• Create QrySopWand and QrySopWsum classes by copying the QrySopAnd class, deleting the parts that support the boolean retrieval models, and adjusting how scores are calculated.
• Modify the QryParser to recognize the #WSUM and #WAND operators. This is not difficult, but it does require a little care. Weights are determined by their position in the query. #WSUM(2 iphone 1 14) is a valid query. Note that Lucene's lexical analyzer may convert some lexical tokens into two query terms, e.g., Lucene turns "near-death" into "near" and "death". Do not do your own lexical processing of terms, because your choices may not match Lucene's. Just allow for the possibility that one token may produce more than one term. The user query #WSUM(2 near-death 1 experience) should be parsed into the query #WSUM(2 near 2 death 1 experience).

Implement the WINDOW operator. There are three main steps.

• Use the QryIopNear class to guide your implementation of a QryIopWindow class.
• The evaluate method is the heart of each QryIopXxx operator. It uses the arguments (which are always QryIopXxx operators that have inverted lists) to produce a new inverted list.
• Modify the QryParser to recognize #WINDOW operators in queries.

That's about it.

FAQ

If you have questions not answered here, see the HW2 FAQ.

Copyright 2024, Carnegie Mellon University.
Updated on February 07, 2024