pyndri

pyndri is a Python interface to the Indri search engine


Keywords
indri, language, models, retrieval, indexing, indri-search-engine, information-retrieval, python, research
License
MIT
Install
pip install pyndri==0.4

Documentation

pyndri

pyndri is a Python interface to the Indri search engine (http://www.lemurproject.org/indri/).

Requirements

During development, we use Python 3.5. Some of the examples require numpy.

Installation

You first need to install Indri before you can install pyndri.

Linux

While there are many different Linux distributions, the following seems to work for most. Below we assume a Debian-based system (tested on Ubuntu and Mint), however, your mileage may vary.

First, install dependencies:

sudo apt install g++ zlib1g-dev python3.5-dev python3-pip
sudo pip3 install setuptools

To install Indri:

wget <url-to-indri-5.11.tar.gz>
tar xzvf indri-5.11.tar.gz
cd indri-5.11
./configure CXX="g++ -D_GLIBCXX_USE_CXX11_ABI=0"
make
sudo make install

Finally, install pyndri:

sudo pip3 install pyndri

macOS

The instructions for macOS as similar to those for Linux. You most likely want to install a modern version of Python, GCC and the development package of zlib using Homebrew or MacPorts.

Windows + Cygwin (experimental)

To get Indri running under Windows with Cygwin, a small adjustment needs to be made to the above instructions. First, you probably want to install the apt-cyg package manager and install a modern version of Python, GCC and the zlib development package.

To install Indri:

wget <url-to-indri-5.11.tar.gz>
tar xzvf indri-5.11.tar.gz
cd indri-5.11
./configure CXX="g++ -D_GNU_SOURCE=1 -D_GLIBCXX_USE_CXX11_ABI=0"
make
sudo make install

Finally, install pyndri:

sudo pip install pyndri

Toolkit examples

Since v0.3, pyndri comes with a few tools to extract index statistics and to generate rankings using common retrieval models.

PyndriStatistics

The following example shows statistics about the New York Times collection, as used in the TREC Common Core 2017 track.

> PyndriStatistics --index /path/to/nyt-index
[INFO]  Arguments: Namespace(index='/path/to/nyt-index', loglevel='INFO')
[INFO]  NUM=1855658
[INFO]  LENGTH_MEDIAN=414.0
[INFO]  LENGTH_MODE=109.0
[INFO]  LENGTH_MEAN=572.1752801432273
[INFO]  LENGTH_MIN=0.0
[INFO]  LENGTH_MAX=36814.0
[INFO]  LENGTH_STD=605.8240048
[INFO]  TOTAL_TERMS=1061761636
[INFO]  UNIQUE_TERMS=1351413

PyndriQuery

The following example runs QLM with Dirichlet smoothing (with mu equal to the average document length) and pseudo-relevance feedback on the TREC Common Core 2017 test collection.

> PyndriQuery \
	--loglevel warning \
	--queries commoncore2017_queries.txt \
	--index /path/to/nyt-index \
	--smoothing_method dirichlet --smoothing_param auto --prf \
	test.run

> trec_eval commoncore2017.qrel test.run-commoncore2017_queries.txt | grep -E "^map\s+"
map                   	all	0.2499

API examples

How to iterate over all documents in a repository:

import pyndri

index = pyndri.Index('/path/to/indri/index')

for document_id in range(index.document_base(), index.maximum_document()):
    print(index.document(document_id))

The above will output pairs of external document identifiers and document terms:

('eUK237655', (877, 2171, 191, 2171))
('eUK956325', (880, 2171, 345, 2171))
('eUK458961', (3566, 1, 2, 3199, 504, 1726, 1, 3595, 1860, 1171, 1527))
('eUK390317', (3228, 2397, 2, 945, 1, 3097, 3, 145, 3769, 2102, 1556, 970, 3959))
('eUK794201', (770, 247, 1686, 3712, 1, 1085, 3, 830, 1445))

How to launch a Indri query to an index and get the identifiers and scores of retrieved documents:

import pyndri

index = pyndri.Index('/path/to/indri/index')

# Queries the index with 'hello world' and returns the first 1000 results.
results = index.query('hello world', results_requested=1000)

for int_document_id, score in results:
    ext_document_id, _ = index.document(int_document_id)
    print(ext_document_id, score)

The above will print document identifiers and language modeling scores:

eUK306804 -8.77414652243
eUK700967 -8.8712247934
eUK437700 -8.88184436222
eUK107263 -8.89119022464
...

The token to term identifier mapping can be extracted as follows:

import pyndri

index = pyndri.Index('/path/to/indri/index')
token2id, id2token, id2df = index.get_dictionary()

id2tf = index.get_term_frequencies()

Citation

If you use pyndri to produce results for your scientific publication, please refer to our ECIR 2017 paper.

@inproceedings{VanGysel2017pyndri,
  title={Pyndri: a Python Interface to the Indri Search Engine},
  author={Van Gysel, Christophe and Kanoulas, Evangelos and de Rijke, Maarten},
  booktitle={ECIR},
  volume={2017},
  year={2017},
  organization={Springer}
}

Frequently Asked Questions

Importing pyndri in Python causes the error Undefined symbol std::__cxx11::basic_string ...

You are using GCC 5 (or above) and this version of the compiler includes new implementations of common types (std::string, etc.). You have to recompile Indri first by setting the _GLIBCXX_USE_CXX11_ABI macro to 0.

make clean
./configure CXX="g++ -D_GLIBCXX_USE_CXX11_ABI=0"
make
sudo make install

Afterwards, recompile pyndri from a clean install.

Importing pyndri in Python causes the error undefined symbol: std::throw_out_of_range_fmt ...

Your Python version was compiled with a different standard library than you used to compile Indri and pyndri with. For example, the Anaconda distribution comes with pre-compiled binaries and its own standard library.

We do not provide support for this, as this is a problem with your Python installation and goes beyond the scope of this project. However, we've identified three possible paths:

  • Re-compile Python yourself from source.
  • Compile Indri and pyndri with the standard library of your Python distribution. This might be difficult, as the headers are often not included in the distribution.
  • Use the Python executables part of your Linux distribution. Be sure to install the development headers (e.g., python3.5-dev using apt-get on Ubuntu).

License

Pyndri is licensed under the MIT license. Please note that Indri is licensed separately. If you modify Pyndri in any way, please link back to this repository.