ntcir-datasearch

Baselines for NTCIR Data Search

Setup

Installation of Poetry (skip this step if Poetry has already been installed)

$ curl -sSL https://raw.githubusercontent.com/python-poetry/poetry/master/get-poetry.py | python

Installation of required Python packages

$ source ~/.bash_profile
$ poetry install

Installation of Java

Since this package uses Anserini, Java 11 and Maven 3.3+ are also required.

Downloading Data Search files

Please visit https://ntcir.datasearch.jp/ and download the test collection, which includes

• data_search_j_collection.jsonl.bz2
• data_search_e_collection.jsonl.bz2
• data_search_j_train_topics.tsv
• data_search_e_train_topics.tsv

These files are expected to be in data directory.

Build

Please run the command below for compiling Java codes of Anserini:

$ poetry run invoke build

If the build successes, you will see a message that looks like

[INFO] ------------------------------------------------------------------------
[INFO] BUILD SUCCESS
[INFO] ------------------------------------------------------------------------
[INFO] Total time:  37.788 s
[INFO] Finished at: 2020-03-06T10:58:14+09:00
[INFO] ------------------------------------------------------------------------

Compiled anserini successfully.

Indexing

Let's index the collection of Japanese statistical data.

preprocess task reads contents from data/data_search_j_collection.jsonl, and produces multiple files in JSONL format.

$ poetry run invoke preprocess ja data/data_search_j_collection.jsonl collections/ja

You can see the generated files by

$ ls collections/ja
collection.00000000.jsonl collection.00000003.jsonl collection.00000006.jsonl collection.00000009.jsonl collection.00000012.jsonl
collection.00000001.jsonl collection.00000004.jsonl collection.00000007.jsonl collection.00000010.jsonl collection.00000013.jsonl
collection.00000002.jsonl collection.00000005.jsonl collection.00000008.jsonl collection.00000011.jsonl

When you process the English statistical data data_search_e_collection.jsonl, ja should be replaced with en.

Then, index task indexes the collection as follows:

$ poetry run invoke index ja collections/ja indices/ja
...
2020-03-06 13:22:53,100 INFO  [main] index.IndexCollection (IndexCollection.java:841) - Total 1,338,402 documents indexed in 00:02:27

Created the index successfully.

You can find the index files at indices/ja.

Search

After the index has been built, search task can produce ranked lists by several search models such as BM25 and LMIR.

$ poetry run invoke search ja indices/ja data/data_search_j_train_topics.tsv results

This will read queries from data/data_search_j_train_topics.tsv, retrieve results from indices/ja, and output them into results directory.

$ ls results/
ja-bm25.accurate.txt ja-bm25.txt          ja-bm25prf+bm25.txt  ja-qld.txt           ja-rm3+bm25.txt      ja-rm3+qld.txt       ja-sdm+bm25.txt      ja-sdm+qld.txt

The output file name is <language>-<search_model>.txt, where

• bm25 : BM25 scoring model
• bm25.accurate : BM25 scoring model
• bm25prf : bm25PRF query expansion model
• qld : query likelihood Dirichlet scoring model
• rm3 : RM3 query expansion model
• sdm : Sequential Dependence Model

Please refer to Anserini for details on search models.

Note that these results are in the TREC format, not the NTCIR format required in the NTCIR-15 Data Search task. You can transform a TREC file into an NTCIR file by ntcirify command, e.g. poetry run invoke ntcirify trec_file.txt ntcir_file.txt.