Skip to content

Latest commit

 

History

History
441 lines (342 loc) · 18.7 KB

runbook-trec2018-h2oloo.md

File metadata and controls

441 lines (342 loc) · 18.7 KB
Raw

TREC 2018 Runbook: "h2oloo" Group

This is the runbook for TREC 2018 submissions by the "h2oloo" group. Note that Anserini (the system) was used by another group ("Anserini") for a completely different set of runs. The h2oloo group participated in the Common Core Track.

Note that this document is specifically a runbook and does not encode regression experiments. Runbooks are designed to help us (i.e., TREC participants) document the steps taken to generate a run. They are primarily designed to make experiments repeatable (i.e., by ourselves), although they might be helpful for others who wish to reproduce our runs.

However, we concede that repeatability of the runs (even by us) is challenging, since the codebase is always evolving, and by the time we add proper documentation, it might be several months later... but we try our best...

Topics and qrels are stored in tools/topics-and-qrels/.

Data Preparation with Anserini

Index Construction

Robust04:

target/appassembler/bin/IndexCollection -collection TrecCollection \
 -generator JsoupGenerator -threads 16 -input /path/to/robust04 \
 -index lucene-index.robust04.pos+docvectors+rawdocs \
 -storePositions -storeDocvectors -storeRawDocs >& log.robust04.pos+docvectors+rawdocs

Robust05:

target/appassembler/bin/IndexCollection -collection TrecCollection \
 -generator JsoupGenerator -threads 16 -input /path/to/robust05 \
 -index lucene-index.robust05.pos+docvectors+rawdocs \
 -storePositions -storeDocvectors -storeRawDocs >& log.robust05.pos+docvectors+rawdocs

Core17:

target/appassembler/bin/IndexCollection -collection NewYorkTimesCollection \
 -generator JsoupGenerator -threads 16 -input /path/to/core17 \
 -index lucene-index.core17.pos+docvectors+rawdocs \
 -storePositions -storeDocvectors -storeRawDocs >& log.core17.pos+docvectors+rawdocs

Core18:

target/appassembler/bin/IndexCollection -collection WashingtonPostCollection \
 -generator WapoGenerator -threads 16 -input /path/to/WashingtonPost.v2/data/ \
 -index lucene-index.core18.pos+docvectors+rawdocs \
 -storePositions -storeDocvectors -storeRawDocs &> log.core18.pos+docvectors+rawdocs

Initial Retrieval

Robust04:

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.robust04.pos+docvectors+rawdocs \
 -topics tools/topics-and-qrels/topics.robust04.301-450.601-700.txt \
 -output run.robust04.bm25+rm3.topics.robust04.301-450.601-700.txt -bm25 -rm3 -hits 10000

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.robust04.pos+docvectors+rawdocs \
 -topics tools/topics-and-qrels/topics.robust04.301-450.601-700.txt \
 -output run.robust04.bm25+ax.topics.robust04.301-450.601-700.txt \
 -bm25 -axiom -rerankCutoff 20 -axiom.deterministic -hits 10000

Robust05:

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.robust05.pos+docvectors+rawdocs \
 -topics tools/topics-and-qrels/topics.robust05.txt \
 -output run.robust05.bm25+rm3.topics.robust05.txt -bm25 -rm3 -hits 10000

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.robust05.pos+docvectors+rawdocs \
 -topics tools/topics-and-qrels/topics.robust05.txt \
 -output run.robust05.bm25+ax.topics.robust05.txt \
 -bm25 -axiom -rerankCutoff 20 -axiom.deterministic

Core17:

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.core17.pos+docvectors+rawdocs \
 -topics tools/topics-and-qrels/topics.core17.txt \
 -output run.core17.bm25+rm3.topics.core17.txt -bm25 -rm3 -hits 10000

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.core17.pos+docvectors+rawdocs \
 -topics tools/topics-and-qrels/topics.core17.txt \
 -output run.core17.bm25+ax.topics.core17.txt \
 -bm25 -axiom -rerankCutoff 20 -axiom.deterministic -hits 10000

Core18:

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.core18.pos+docvectors+rawdocs -bm25 \
 -topics tools/topics-and-qrels/topics.core18.txt \
 -output run.core18.bm25+rm3.topics.core18.txt -rm3 -hits 10000

target/appassembler/bin/SearchCollection -topicreader Trec \
 -index lucene-index.core18.pos+docvectors+rawdocs -bm25 \
 -topics tools/topics-and-qrels/topics.core18.txt \
 -output run.core18.bm25+ax.topics.core18.txt -axiom \
 -rerankCutoff 20 -axiom.deterministic -hits 10000

Extraction of TF-IDF vectors

Robust04:

target/appassembler/bin/IndexUtils -index lucene-index.robust04.pos+docvectors+rawdocs \
 -dumpAllDocids NONE

target/appassembler/bin/IndexUtils -index lucene-index.robust04.pos+docvectors+rawdocs \
 -dumpDocVectors lucene-index.robust04.pos+docvectors+rawdocs.allDocids.txt \
 -docVectorWeight TF_IDF

Robust05:

target/appassembler/bin/IndexUtils -index lucene-index.robust05.pos+docvectors+rawdocs \
 -dumpAllDocids NONE

target/appassembler/bin/IndexUtils -index lucene-index.robust05.pos+docvectors+rawdocs \
 -dumpDocVectors lucene-index.robust05.pos+docvectors+rawdocs.allDocids.txt \
 -docVectorWeight TF_IDF

Core17:

target/appassembler/bin/IndexUtils -index lucene-index.core17.pos+docvectors+rawdocs \
 -dumpAllDocids NONE

target/appassembler/bin/IndexUtils -index lucene-index.core17.pos+docvectors+rawdocs \
 -dumpDocVectors lucene-index.core17.pos+docvectors+rawdocs.allDocids.txt \
 -docVectorWeight TF_IDF

Core18:

target/appassembler/bin/IndexUtils -index lucene-index.core18.pos+docvectors+rawdocs \
 -dumpAllDocids NONE

target/appassembler/bin/IndexUtils -index lucene-index.core18.pos+docvectors+rawdocs \
 -dumpDocVectors lucene-index.core18.pos+docvectors+rawdocs.allDocids.txt \
 -docVectorWeight TF_IDF

Classifier Training/Inference with Scikit-Learn

Now we are done with Anserini part; everything below is in Python.

Create the following folder structure and populate with files as follows:

root
 |
 |--+ main_folder
 |   |-- train
 |   |-- test
 |
 |--+ src
 |  |--+ main_folder
 |     |--+ train
 |     |  |--+ tfidf
 |     |  |  |-- Robust04_tfidf.tar.gz
 |     |  |  |-- Robust05_tfidf.tar.gz
 |     |  |  |-- Core17_tfidf.tar.gz
 |     |  |
 |     |  +--+ rank
 |     |     |-- Robust04_qrels.txt
 |     |     |-- Robust05_qrels.txt
 |     |     |-- Core17_qrels.txt
 |     |
 |     |--+ test
 |        |-- Core18_tfidf.tar.gz
 |        |-- Core18_rerank_ax.txt
 |        |-- Core18_rerank_rm3.txt
 |
 |--+ script
    |-- build_train.py
    |-- build_test.py
    |-- train.py
    |-- submission.py
    |-- utils.py

The source Python scripts are in src/main/python/trec2018/h2oloo-core/.

Building Training Data

Run the following command to build the training data:

python script/build_train.py --tfidf-folder src/main_folder/train/tfidf/ \
 --qrels-folder src/main_folder/train/rank/ --output-folder main_folder/train/

The command performs the following:

  • Building vocabulary. Training data consists of documents in qrels from Robust04, Robust05, and Core17. Vocabulary contains all words in training data.
  • Building and Dumping docid index and vocabulary index. Each document and each vocabulary in training data are indexed for the ease of building feature matrix in next step. They are dumped in main_folder/train/.
  • Building features and labels. Features are the TF-IDF values with shape (n_docs, n_vocabs) topic-wise, and labels are the qrels value (changing all 2s to 1s). Features are stored using scipy.sparse_matrix
  • Dump features and labels. Features are dumped as .npz file and labels are as .npy file for each topic. These files are named with the topic id, e.g., 321.npz is the feature file for topic 321 and 321.npy is the label file for topic 321. They are dumped in main_folder/train/features folder.

Building Test Data

Run the following command to build test data:

python script/build_test.py --tfidf-file src/main_folder/test/Core18_tfidf.tar.gz \
 --rank-file src/main_folder/test/Core18_rerank_[rm3|ax].txt \
 --output-folder main_folder/test/ --vocab-folder main_folder/train/

The command performs the following:

  • Building and Dumping docid index. Test data consists of top 10000 documents generated by BM25+RM3 or BM25+Axiom for each topic. Each document in test data is indexed. The docid index file is dumped in main_folder/test.
  • Building features. Features are TF-IDF values with shape (n_docs, n_vocabs), where vocabulary is build in the above step.
  • Dump features. A single .npz file was dumped at this step. In the dev step, we can use docid index to find the features for any document. The docid index file is dumped in main_folder/test.

Training and Generating Classifier Results

Run the following command to train classifiers and run inference on Core18 documents topic-wise to rerank the documents.

python script/train.py --train-folder main_folder/train/ \
 --test-folder main_folder/test/ \
 --rank-file src/main_folder/test/Core18_rerank_[rm3|ax].txt \
 --classifier clf --output-folder main_folder/output/clf

Where clf is one of the following:

  • lr1
  • lr2
  • svm
  • sgdr (SGD Regressor)
  • sgdc (SGD Classifier)
  • lgb (lightGBM)
  • ridge

The result will be in main_folder/output/CLASSIFIER folder, where CLASSIFIER is the classifier used for training.

Note that only one of rm3 or axiom results can be generated at one time and fed into next step (because the output name are fixed).

Generating Submission Files

Run the following script to apply classifiers (either a single classifier or an ensemble) to generate the submissions. The argument --ensemble can be {1, 3, 7}:

  • with 1, LR2 is used,
  • with 3, LR2+SVM+LGB is used, and
  • with 7, all classifiers are used.

BM25+Axiom

  • LR2 + SVM + LGB on BM25+axiom ranking, interpolation weight 0.6 (Priority 1)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_ax.txt \
 --clf-folder main_folder/output --ratio 0.6 --ensemble 3 \
 --runtag h2oloo_e3ax0.6 --output core18.h2oloo_e3ax0.6.txt
  • LR2 on BM25+axiom ranking, interpolation weight 0.6 (Priority 3)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_ax.txt \
 --clf-folder main_folder/output --ratio 0.6 --ensemble 1 \
 --runtag h2oloo_LRax0.6 --output core18.h2oloo_LRax0.6.txt
  • LR2 + SVM + LGB on BM25+axiom ranking, interpolation weight 0.7 (Priority 2)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_ax.txt \
 --clf-folder main_folder/output --ratio 0.7 --ensemble 3 \
 --runtag h2oloo_e3ax0.7 --output core18.h2oloo_e3ax0.7.txt
  • All classifiers on BM25+axiom ranking, interpolation weight 0.6 (Priority 5)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_ax.txt \
 --clf-folder main_folder/output --ratio 0.6 --ensemble 7 \
 --runtag h2oloo_e7ax0.6 --output core18.h2oloo_e7ax0.6.txt
  • All classifiers on BM25+axiom ranking, interpolation weight 0.7 (Priority 6)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_ax.txt \
 --clf-folder main_folder/output --ratio 0.7 --ensemble 7 \
 --runtag h2oloo_e7ax0.7 --output core18.h2oloo_e7ax0.7.txt

BM25+RM3

  • LR2 + SVM + LGB on BM25+RM3 ranking, interpolation weight 0.6 (Priority 4)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_rm3.txt \
 --clf-folder main_folder/output --ratio 0.6 --ensemble 3 \
 --runtag h2oloo_e3rm30.6 --output core18.h2oloo_e3rm30.6.txt
  • All classifiers on BM25+RM3 ranking, interpolation weight 0.6 (Priority 7)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_rm3.txt \
 --clf-folder main_folder/output --ratio 0.6 --ensemble 7 \
 --runtag h2oloo_e7rm30.6 --output core18.h2oloo_e7rm30.6.txt
  • All classifiers on BM25+RM3 ranking, interpolation weight 0.7 (Priority 8)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_rm3.txt \
 --clf-folder main_folder/output --ratio 0.7 --ensemble 7 \
 --runtag h2oloo_e7rm30.7 --output core18.h2oloo_e7rm30.7.txt
  • LR2 on BM25+RM3 ranking, interpolation weight 0.6 (Priority 10)
python script/submission.py --rank-file src/main_folder/test/Core18_rerank_rm3.txt \
 --clf-folder main_folder/output --ratio 0.6 --ensemble 1 \
 --runtag h2oloo_LRrm0.6 --output core18.h2oloo_LRrm0.6.txt

Effectiveness

Run trec_eval to evaluate effectiveness. On the runs based on Axiom:

../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_LRax0.6.txt
../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_e3ax0.6.txt
../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_e3ax0.7.txt
../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_e7ax0.6.txt
../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_e7ax0.7.txt

On the runs based on RM3:

../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_LRrm0.6.txt
../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_e3rm30.6.txt
../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_e7rm30.6.txt
../eval/trec_eval.9.0.4/trec_eval -c -M1000 -m map -m ndcg -m P.10 ../tools/topics-and-qrels/qrels.core18.txt core18.h2oloo_e7rm30.7.txt

We have confirmed that as of commit acf4c872f6bf36756ba972dbddd8fcefcfb2a648 (Sat Dec 15 11:41:59 2018 -0500), the above commands work. For reference, this was the conda environment (setting up the environment with conda install -c conda-forge lightgbm):

$ conda list
# packages in environment at /anaconda3/envs/python36:
#
# Name                    Version                   Build  Channel
blas                      1.0                         mkl  
bzip2                     1.0.6                         1    conda-forge
ca-certificates           2018.11.29           ha4d7672_0    conda-forge
certifi                   2018.11.29            py36_1000    conda-forge
clangdev                  4.0.0                 default_0    conda-forge
icu                       58.2                 hfc679d8_0    conda-forge
intel-openmp              2019.1                      144  
libcxx                    4.0.1                hcfea43d_1  
libcxxabi                 4.0.1                hcfea43d_1  
libedit                   3.1.20170329         hb402a30_2  
libffi                    3.2.1                h475c297_4  
libgfortran               3.0.1                h93005f0_2  
libiconv                  1.15                 h470a237_3    conda-forge
libxml2                   2.9.8                h422b904_5    conda-forge
lightgbm                  2.2.1            py36hfc679d8_0    conda-forge
llvmdev                   4.0.0                 default_0    conda-forge
mkl                       2019.1                      144  
mkl_fft                   1.0.10                   py36_0    conda-forge
mkl_random                1.0.2                    py36_0    conda-forge
ncurses                   6.1                  h0a44026_1  
numpy                     1.15.4           py36hacdab7b_0  
numpy-base                1.15.4           py36h6575580_0  
openmp                    4.0.0                         1    conda-forge
openssl                   1.0.2p               h470a237_1    conda-forge
pip                       18.1                     py36_0  
python                    3.6.6                h5001a0f_0    conda-forge
readline                  7.0                  h1de35cc_5  
scikit-learn              0.20.1           py36h27c97d8_0  
scipy                     1.1.0            py36h1410ff5_2  
setuptools                40.6.2                   py36_0  
sqlite                    3.25.3               ha441bb4_0  
tk                        8.6.8                ha441bb4_0  
wheel                     0.32.3                   py36_0  
xz                        5.2.4                h1de35cc_4  
zlib                      1.2.11               h1de35cc_3

However, the effectiveness of the runs differs from the effectiveness of our official TREC runs submitted to NIST.

In the table below, we compare the effectiveness of our submitted runs, marked with * (e.g., AP*), and the effectiveness of the generated runs at the commit point referenced above, marked with + (e.g., AP+).

Metric AP* AP+ NDCG* NDCG+ P10* P10+
h2oloo_LRax0.6 0.3227 0.3343 0.6123 0.6297 0.5800 0.5780
h2oloo_e3ax0.6 (h2oloo_enax0.6) 0.3341 0.3507 0.6233 0.6450 0.5860 0.6060
h2oloo_e3ax0.7 (h2oloo_enax0.7) 0.3351 0.3477 0.6218 0.6463 0.5920 0.6100
h2oloo_e7ax0.6 0.3310 0.3464 0.6215 0.6433 0.5840 0.6080
h2oloo_e7ax0.7 0.3274 0.3455 0.6209 0.6434 0.5880 0.6020
h2oloo_LRrm0.6 (h2oloo_LR2_rm3) 0.3273 0.3539 0.6113 0.6471 0.5800 0.6040
h2oloo_e3rm30.6 (h2oloo_enrm30.6) 0.3382 0.3652 0.6193 0.6533 0.5920 0.6140
h2oloo_e7rm30.6 0.3333 0.3620 0.6143 0.6510 0.5820 0.6100
h2oloo_e7rm30.7 0.3361 0.3609 0.6177 0.6498 0.5940 0.6040

There is a one additional run h2oloo_LR2AX0.6 submitted to TREC that came from another code base, which we do not include here.

Update: As of commit e71df7aee42c7776a63b9845600a4075632fa11c Tue Dec 18 07:45:30 2018 -0500, Anserini was upgraded to Lucene 7. The results are different from those with Lucene 6 above. Differences are as follows between code running at the above commit point (L7) vs. previous results (L6):

Metric AP* AP+ L6 AP+ L7
h2oloo_LRax0.6 0.3227 0.3343 0.3396
h2oloo_e3ax0.6 (h2oloo_enax0.6) 0.3341 0.3507 0.3563
h2oloo_e3ax0.7 (h2oloo_enax0.7) 0.3351 0.3477 0.3526
h2oloo_e7ax0.6 0.3310 0.3464 0.3504
h2oloo_e7ax0.7 0.3274 0.3455 0.3490
h2oloo_LRrm0.6 (h2oloo_LR2_rm3) 0.3273 0.3539 0.3569
h2oloo_e3rm30.6 (h2oloo_enrm30.6) 0.3382 0.3652 0.3670
h2oloo_e7rm30.6 0.3333 0.3620 0.3639
h2oloo_e7rm30.7 0.3361 0.3609 0.3647