L. Gallagher, A. Mallia, J. S Culpepper, T. Suel, and B. Barla Cambazoglu. 2020. Feature Extraction for Large-Scale Text Collections. In Proc. CIKM. DOI: https://doi.org/10.1145/3340531.3412773

This repository contains scripts to build the dataset, and reproduce the experiments from the paper Feature Extraction for Large-Scale Text Collections from the CIKM 2020 Resource Track.

The dataset is available for download at the links below:

• Release 1.0.1 - cikm20ltr-1.0.1 (sha1 cca713f3d331921f4d5d3093832d5f182da79c25)

The following environment configuration was used to build the dataset and run the experiments. We assume you have a working conda installation (recommended).

Clone this repo and setup Conda environment:

git clone https://github.com/ten-blue-links/cikm20
cd cikm20
git submodule update --init --recursive --depth 1
conda env create -f env.yml
conda activate cikm20fxt
./src/sh/lgbm.sh
pip install -r requirements.txt

The following details the prerequisites and steps to configure and run the build scripts.

• Indri index of ClueWeb09B (example config)
• ~350GiB RAM
• ~300GiB disk space
• Webgraph data ClueWeb09_WG_50m.graph-txt.gz and ClueB-ID-DOCNO.txt.tar.gz. Once downloaded decompress the ClueB-ID-DOCNO.txt.tar.gz:
  • ClueWeb09B_WG_50m.graph-txt.gz leave this as is.
  • ClueB-ID-DOCNO.txt.tar.gz decompress to ClueB-ID-DOCNO.txt.
• The gradle build system was used for the AlexaRank data
• GCC 8.x (not tested with Clang)
• Boost (tested with 1.65.1)
• Cmake 3.x

1. Copy configuration template: cp config/dataset.dist config/dataset
2. Edit config/dataset and configure the following variables:
   • INDRI_INDEX_PATH - path to existing ClueWeb09B Indri index (example config)
   • FXT_INDEX_PATH - path where the Fxt index will be created
   • BOOST_INCLUDE_PATH - path to Boost headers
   • BOOST_LIBRARY_PATH - path to Boost libraries
   • INDRI_INCLUDE_PATH - path to Indri headers
   • INDRI_LIBRARY_PATH - path to Indri libraries
   • WEBGRAPH_PATH - path to ClueWeb09_WG_50m.graph-txt.gz (gzipped)
   • GRAPHPAIRS_PATH - path to ClueB-ID-DOCNO.txt (decompressed)
3. Run ./src/dataset/main.sh (build may take ~32 hours)
4. Dataset files build/cikm20ltr

The snapshot for the AlexaRank data is from 2010. This was the temporally closest working snapshot to Jan-Feb 2009 for ClueWeb09B.

The term reproduce is defined as per the ACM artifacts policy. Note the definitions for the terms replicate and reproduce were recently swapped around (Aug 2020).

1. Copy configuration template: cp config/experiment.dist config/experiment
   1. If the dataset files are in a different location than the default build/cikm20ltr edit config/experiment and set DATASETD to the correct path
2. Run ./src/experiment/main.sh
3. cat the results: for i in build/result/wt??/test/eval/*.txt; do echo $i; cat $i; done
4. TREC run files build/result/wt??/test/run

The experiment scripts should be able to reproduce the following results: