Introduction

This is a information retrival system used to index the trec-cds data, and do the training and search work. The system is mainly based on Lucene, and the project has contained the jar package of Lucene and some other tools.

Environment

• java
• Linux
• shell

System Structure

• source files
  • TransOfFormat : to transform to trec format
  • indexing.Indexer : to index files
  • trec.SearchTopics
    • train the BM25 params
    • search topics
• lib
  • lucene jar packages
  • dom4j
• bin
  • java compiled class files
• data directory
  • pmc-text : store the extracted source trec files
  • index : store index files
  • sample_evavl.pl : used to evaluate
  • others
    • topics
    • qrels
    • some tmp files
• executable shell script
  • compile.sh : to compile java source files
  • parse.sh : to parse the file to trec format
  • indexing.sh : to index files
  • search.sh : to search the topics and generate the result file

Realization

• transform files format
  • use TransOfFormat tools to transform the files to regular trec files
• index
  • use lucene to index the files, and save index to ../data/index
• train params
  • traverse b from [b1, b2] and k1 from [k', k'']
  • search the topics14.xml, generate a tmp file
  • use perl sample_eval.pl to evaluate the tmp file, and compare the NDCG value
  • choose the best k1 and b
• search topics15
  • use the best k1 and b to do the search

Main java class

• tool TransOfFormat
• indexing.Indexer.java
• trec.SearchTopics.java

Run

one-key run is

bash run.sh

and can be divided to several parts:

• compile
  • bash compile.sh
• parse
  • bash parse.sh
• index
  • bash index.sh
• search
  • bash search.sh

Technologies & Principles

• BM25
  • use lucene BM25Similarities
  • the params k1 and b are trained with the qrels2014

searcher.setSimilarity(new BM25Similarity(pa.k1, pa.b));

Experiment Procedure

• preprocess
  • extract the downloaded compressed files to ../data/pmc-text
  • use the tool TransFormat to transform the files format
• index
  • read the trec format files
  • get the DOCID information, and get the text information
  • use Lucene index writer to index
• train
  • traverse each k1 and b and do search 2014 queries
  • compare result choose the best k1 and b
  • from the following code, it is obvious that the training set is topics2014.xml

String queries14 = "../data/topics2014.xml";
String queries15 = "../data/topics2015A.xml";
String indexFolder = "../data/index";
String qrelsFile14 = "../data/qrels2014.txt";
String queryResult15A = "../data/queryResult15A.txt";
Param pa = trainBM25Param(queries14, indexFolder, qrelsFile14);

• search
  • search 2015A queries with the best k1 and b